Rather than relying on throttling, batching, or conflation to reduce the flow of updates, Rithmic provides un-throttled, un-aggregated market data, delivering exchange updates as they are received.

For active futures traders—particularly those using the DOM, order flow, or other market-microstructure tools—this can provide a more granular, real-time view of changes in price, volume, and displayed liquidity.

• Rithmic, LLC is transmitting market data since 2006

• Rithmic offers the complete trading technology stack: Institutional-grade market data, execution and risk management.

• Rithmic is FCM-neutral and works with all major futures brokers, funding evaluators (prop firms), and FCMs.

When traders think about execution quality, the focus often falls on the broker or trading platform. Yet the quality of the information reaching the trader is another important part of the equation.

A trader can only make decisions based on the market information available to them.

This is particularly relevant for short-term futures traders, whose decisions may depend not only on where the market is trading, but on how liquidity is appearing, disappearing, and shifting across the order book.

That is where the distinction between aggregated and un-aggregated market data becomes important.

Aggregated vs. Un-Aggregated Data

Aggregation is, essentially, a way of simplifying market information.

Multiple orders at the same price level can be combined into a single displayed quantity. The resulting order book is easier to process and interpret, but some of the underlying detail is no longer visible.

An un-aggregated feed preserves more of that detail. Instead of presenting only a consolidated quantity at a price level, it can provide greater visibility into the individual orders and quantities that make up the displayed liquidity.

For a trader who primarily watches charts, that distinction may have little practical significance.

For someone trading from the DOM, however, it can be much more relevant.

The order book is not simply a list of prices. It is a constantly changing representation of displayed liquidity. Traders who study it may look for signs of aggressive buying and selling, resting liquidity, orders being added or pulled, absorption, and imbalances.

The more detail a trader has, the more context they may have when interpreting those changes.

That does not mean that un-aggregated data automatically produces better trading decisions. More information is not the same as better information. The potential advantage comes from what a trader can interpret from that information.

Reading the Market Through the DOM: Consider a simplified example.

Suppose a futures trader sees 500 contracts displayed at a particular price level. In an aggregated order book, that may simply appear as one block of 500 contracts.

A more granular view could reveal that the 500 contracts are made up of a number of individual orders, each with different quantities.

That distinction can provide additional context.

Is the liquidity concentrated in a single large order? Is it spread across many smaller orders? Is liquidity being added as the market approaches the level, or pulled away before price reaches it?

For an active trader studying market microstructure, these are meaningful questions.

The order book, however, should never be treated as a crystal ball. Displayed liquidity is not a guarantee of future execution or price direction. Orders can be modified or canceled, and the presence of liquidity at a given price does not necessarily mean that it will remain there when the market reaches that level.

The value of granular data therefore lies less in predicting the future than in giving traders a more detailed picture of the present.

Rithmic and the Active-Trader Use Case: Reading the DOM

• Create a hypothetical futures-market scenario.

• Show how an aggregated display might indicate substantial liquidity at a particular price.

• Then explain how more granular information could reveal that the apparent liquidity consists of multiple individual orders with different sizes.

• Discuss how an active trader might use that additional context.

• Stress that order-book information represents displayed liquidity, not a guaranteed prediction of future price direction.

There is also a trade-off.

Higher-resolution market data means more information to process. That can require greater bandwidth, more processing capacity, and software capable of handling a much higher volume of market events.

It can also place greater demands on the trader.

A highly detailed order book is useful only if the trader knows what to do with the information. Watching every order appear and disappear can easily become a distraction rather than an advantage.

In other words, data granularity does not create an information advantage by itself.

The advantage, if there is one, comes from the trader’s ability to distinguish meaningful market behavior from noise.

For some strategies, additional detail may be essential. For others, it may simply create more complexity.

Questions Active Traders Should Ask About a Futures Data Feed

• Is the feed aggregated or un-aggregated?

• Which exchanges and contracts are supported?

• How much market depth is available?

• How quickly is market data delivered?

• How does the platform represent the data?

• Can the trader access DOM and order-flow information?

• What happens during periods of extreme volume?

• What are the connectivity and failover arrangements?

• Are there exchange or platform-specific data fees?

• Does the trader’s strategy actually require this level of detail?

Where Rithmic Fits Into the Picture

Rithmic is best understood as one component of a larger trading technology stack.

At a high level, that stack looks something like this:

Every layer plays a role.

The exchange generates the market events. The market-data infrastructure delivers them. The connection and feed determine how that information reaches the trading environment. The platform then processes and displays it, while the trader ultimately decides how to interpret and act on it.

That is why comparing feeds solely on a headline latency figure can be misleading. A trader’s experience depends on the entire chain, from the exchange to the screen—and, ultimately, to the order sent back into the market.

If you trade primarily from charts, granular order-book data may offer limited additional value but If you trade from the DOM, the level of market-data detail becomes much more relevant.

If you use order-flow analytics, it is worth understanding exactly how your platform receives, processes, and represents the underlying feed.

If you scalp heavily, market-data quality, latency, reliability, and execution infrastructure deserve closer scrutiny.

And if you trade systematically, the important questions may be different again: message rates, API characteristics, reliability, data handling, and the ability of your infrastructure to process the volume of information generated by the market.

Market data is not merely a stream of prices. It is the raw information from which a trader builds a view of the market.

How Rithmic’s Exchange Simulator Works

Rithmic has spent more than 20 years building and refining a simulator designed to behave as closely as possible to the live market.

Rithmic's paper trading system connects to live exchange data from CME Group and Nasdaq. Real quotes, real order flow — simulated fills, zero capital at risk.

Live market data from CME Group and Nasdaq

All order types: market, limit, stop, bracket, OCO

Server-side risk rules enforced — nearly identical to live accounts

Full support for over a dozen third-party platforms

R | Trader Pro™ desktop platform included

https://www.rithmic.com/products/exchange-simulator

The goal is simple: what you learn in simulation should translate to live trading. That means fills need to reflect the order book, execution needs to follow the same path, and the environment needs to behave like the one you’ll encounter when real money is on the line. Here’s how it works:

• Fills are driven by the order book—not simply the last traded price.
  Our simulator continuously evaluates reported trades and changes in the order book. When the market moves in a way that would have resulted in your order being executed at the exchange, the simulator generates a fill. If the available liquidity is insufficient to fill your entire order at one price, the order can be partially filled across multiple price levels based on the liquidity available in the book. That is how execution works in the live market, and our simulator is designed to reflect it.

• Your order is routed through the trading infrastructure.
  Simulated orders don’t simply get filled locally inside your trading platform. They travel through the same routing infrastructure and execution logic used for live orders, including the effects of network latency. Local fills may look cleaner, but they would not accurately represent the process of getting an order to an exchange and having it interact with the market.

• Your order has a place in the queue.
  Execution is determined by the available bid and ask quantities at each price—not by assuming that a trade at your price automatically means your order was filled. Your order has to interact with the liquidity available ahead of it, just as it would at the exchange. This is important because it prevents artificially optimistic fills that can occur when a simulator relies on price alone.

• The trading environment is the same; the destination is different.
  The simulator operates on the same enterprise-level software infrastructure used by Rithmic’s live trading environment. You receive the same market data feeds and use the same core risk management, order handling, routing, and third-party integration infrastructure. The primary difference is that simulated orders are not sent to the live exchange for execution.

Why this matters

Simulation is most useful when it teaches you to respect the realities of market execution: spread, liquidity, queue position, latency, and available size.

So when a fill in simulation looks unexpected, start by looking at the market rather than assuming the simulator simply “filled you wrong.” Check the spread, the liquidity available at your price, and how the order book moved around the time of the execution.

A thin book, a wide spread, or limited available quantity can produce fills that feel like slippage—but those same conditions can affect execution in the live market.

The purpose of the simulator isn’t to make trading look easier. It’s to make the experience more realistic, so the habits you develop in simulation are habits you can carry into live trading.

CME Group is bringing Single Stock Futures (SSFs) back to the market on July 27, introducing both standard and Micro-sized contracts to provide traders and investors with a flexible, capital-efficient way to gain exposure to individual U.S. stocks.

For hedge funds, proprietary trading firms, and active investors, SSFs offer a centrally cleared, standardized alternative for expressing directional views, implementing relative value strategies, and optimizing financing compared with traditional stock trading.

Press Release by CME Group

Alphabet, Amazon, Apple, Meta, Nvidia and SpaceX among listed firms

Unlike purchasing shares outright, Single Stock Futures enable market participants to speculate on or hedge the price movement of individual companies while using significantly less capital through margin.

With an initial minimum margin requirement of 15%, traders can control the equivalent exposure of 100 shares of stock with a fraction of the capital typically required to purchase those shares outright.

What Are Single Stock Futures (SSFs)?

Single Stock Futures are standardized futures contracts whose value is derived from the price movement of an individual publicly traded company.

PDF (Fact Card)

Each contract:

• Represents 100 shares of the underlying stock.

• Is cash-settled, meaning no physical shares are delivered at expiration.

• Trades on a centrally cleared exchange, reducing counterparty risk.

• Can be used for both speculationand hedging.

Pending the completion of regulatory review and final approval, CME Group plans to launch these contracts on more than 50 of the largest U.S. stocks.

Tactical Trading

Single Stock Futures provide several advantages for active market participants:

• Capital efficiency through lower margin requirements.

• Leverage without purchasing the underlying shares.

• Risk management by hedging existing equity portfolios.

• No physical share delivery, simplifying settlement.

• Standardized contracts backed by CME Group's clearing infrastructure.

• Potential financing advantagescompared with traditional margin trading.

• Guide: Corporate Action Adjustments [CME Group May 26, 2026]

These features make SSFs particularly attractive for relative value strategies, portfolio hedging, and tactical trading.

Before trading Single Stock Futures, it'simportant to understand the mechanics of futures markets, including:

• What a futures contract is.

• Daily mark-to-market settlement.

• Initial and maintenance margin requirements.

• Contract expiration and settlement.

• The impact of leverage on bothprofits and losses.

If you're already familiar with equity index or commodity futures, transitioning to Single Stock Futures is relatively straightforward, as they operate under the same core principles.

Although they may seem new, Single Stock Futures have been around for decades

CME first introduced them in 2002, offering investors another way to gain exposure to individual equities.

While this launch feels new, the product itself has history.

Despite the product's potential, liquidity remained limited and market participation never reached critical mass. As competing products gained traction and exchanges consolidated, Single Stock Futures gradually faded from the spotlight.

Fast forward to 2026, and the environment looks very different.

Demand for capital-efficient tradingproducts has surged, options markets have grown rapidly, andinvestors are more comfortable than ever using derivatives as partof everyday portfolio management.

That shift makes the return of Single Stock Futures especially timely.

While the product showed promise, adoption remained limited due to fragmented liquidity and competition from specialized exchanges such as OneChicago. Following OneChicago's closure, interest in the product gradually faded.

Why They're Returning Now

The market has evolved significantly since the early 2000s.

Today's investors are increasingly seeking more capital-efficient ways to trade, hedge concentrated stock positions, and manage portfolio risk. The 2026 relaunch reflects that demand, combining modern market infrastructure with broader institutionaland retail interest in listed derivatives.

With standard and Micro contract sizes, CME's new offering aims to make Single Stock Futures accessible to a wider range of market participants—from sophisticated retail traders to institutional investors.

The return of CME Group's Single Stock Futures offers a transparent, regulated, and centrally cleared alternative to traditional equity trading.

As with any leveraged product, education and disciplined riskmanagement remain essential. Traders should fully understand marginrequirements, daily settlement, and the risks associated with leveragebefore incorporating Single Stock Futures into their trading strategy.

As with any leveraged instrument, success begins with education. Understanding margin requirements, daily settlement, and risk management is essential before trading Single Stock Futures—but for preparedinvestors, they may become a valuable addition to the trading toolbox.

Where can I trade Single Stock futures?

• Real Trading: Single Stock futures will be available to trade via brokerage firms or your FCM, similar to other Equity product

  

• Prop Trading. Which prop firms will be first to offer these to funded traders?

• This week: Tradeify Futures

  

TradFi Futures VS Perpetuals

CME is offering traditional futures (quarterly expiry, financial settlement, 23-hour/5-day Globex schedule with a daily maintenance break and full weekend closure).

• Hyperliquid offers perpetuals (no expiry, continuous funding-rate anchoring to the oracle, true 24/7 including weekends and holidays). Traders who want to hold directional exposure indefinitely, avoid roll costs/basis decay, or trade outside U.S. equity hours still prefer the perp structure. CME’s product mainly serves institutions and retail who already clear through FCMs and want capital-efficient, regulated exposure during the week; it does not fully substitute for Hyperliquid’s always-on, self-custodied, no-KYC version.

• Demand validation and “larger pie” effect. Hyperliquid already demonstrated strong product-market fit by running high open interest and volume on single-name and pre-IPO equity perps (SpaceX alone has seen hundreds of millions in OI at peaks) well before CME’s announcement. Recent data shows real-world assets (including single stocks) have overtaken pure crypto as a majority of Hyperliquid’s volume in some weeks. CME entering the same names signals that leveraged single-stock derivatives have genuine, growing demand beyond crypto-native users. That awareness can pull more capital and attention into the broader category, with a portion still flowing to the permissionless venue for its unique features (global access, higher effective leverage for some, seamless crypto-collateral settlement).

• User-base and accessibility gaps remain. CME requires brokerage accounts, KYC, and minimum capital typical of futures trading. Hyperliquid has zero minimums, wallet-based access, and serves non-U.S. or privacy-oriented traders who cannot or will not use regulated U.S. venues. Even if CME captures the bulk of institutional hedging flow, Hyperliquid retains the retail/crypto/global niche—and can benefit if some less-sophisticated or weekend-active flow migrates or coexists. Historical precedent in other markets (e.g., crypto perps vs. traditional futures) shows both can grow simultaneously when the overall category expands.

• Narrative and competitive positioning. Success by CME actually reinforces Hyperliquid’s longer-term story as the leading on-chain venue for RWA perpetuals that can take share from traditional exchanges over time. Failure by CME would simply leave the extended-hours single-name demand more exclusively with platforms like Hyperliquid. Either outcome is less damaging than a pure zero-sum interpretation suggests. Regulatory tension (CME/ICE have previously pushed for tighter scrutiny of Hyperliquid) is real, but product traction and volume leadership on the on-chain side continue to build independently.

The SKHYNIX crash

On July 27 at 23:01 UTC, SKHYNIX mark price dropped from $1,127.9 to $917.25. This print was based on an executed trade which was relayed by multiple independent data providers. The XYZ oracle was live in external pricing and tracking that venue, which serves as the primary Korean pre-market venue. The oracle system worked as intended according to its specification.

While the system performed as designed, users are understandably upset about liquidations that were triggered. Market integrity is among Trade[XYZ]'s core values, and we understand the trust of our users has been, and will continue to be, earned.

With these values in mind, we have decided to cover liquidation losses attributable to this anomaly. Eligibility requirements will be conveyed soon, and distributions are anticipated to be completed in the coming days. This is a one-time discretionary decision and is not a guarantee of similar future action.

Going forward, our pricing systems will be further improved to handle tail events. We are accelerating our review of how prices are formed, including revisiting assumptions on external venues, and taking into consideration the price formation which occurs on our own orderbooks, which carry increasingly meaningful depth and signal in relation to external sources.

We will share further information soon.

https://trade.xyz

⚡️

🟩 SP5 > Nvidia > Latest >

​​

☑️ #272 Nov 9, 2025

[Topic: Dysprosium] US curbs block Nvidia’s best AI chips from China (requiring export licenses for even downgraded H20s). China counters by weaponizing its rare-earth chokehold, potentially spiking costs or halting advanced GPU output worldwide

grok.com (Expert) > [Sources not verified. Do your own research] TSMC needs export licenses from China to make Nvidia chips. China produces 99% of Dysprosium used in Nvidia chips. Please, could you offer a concise POV about it?

China processes ~99% of global dysprosium. This heavy rare earth element is now critical for ultrapure capacitors in advanced semiconductors (including Nvidia’s AI GPUs), due to its heat-resistant properties in high-performance chips.

TSMC doesn’t directly need Chinese licenses “to make” Nvidia chips—manufacturing occurs in Taiwan—but recent Chinese export controls (effective November 8, 2025, on dysprosium and other rare earths) mean global suppliers require Beijing’s approval to ship these materials outside China. Without them, TSMC’s supply chain for cutting-edge nodes could face disruptions, indirectly giving China veto power over Nvidia/TSMC production.

+ Related content:

• @thevalueist (10/11/25): $NVDA $AMD $ASML $TSM The Bloomberg chart underscores that China controls a near-monopoly share of refined output for 7 heavy rare earths that are disproportionately used to raise magnet coercivity and temperature tolerance—dysprosium, terbium, gadolinium, samarium, lutetium, yttrium, and scandium—while the U.S. has 0% refined production for these elements. On magnets, China’s market power is even greater at the conversion step: recent IEA data and industry trackers place China’s share of rare-earth magnet output around 85–90%, reflecting choke points at separation and sintered NdFeB magnet manufacturing. This concentration makes any licensing regime or port-of-exit friction function like a quantity restraint on precision motion systems and magnetron assemblies across semiconductor equipment, even when mining tonnage is not the constraint. The combination of functional dominance in magnet manufacturing and zero U.S. refined output for these 7 elements is exactly the profile that gives Beijing maximal leverage over high-spec magnets and sputtering target alloys used in fabs.

  China’s latest rules move the rare earths dispute from symbolic export signaling to a compliance-heavy system with extraterritorial reach. Beijing now requires licenses not just for Chinese-origin exports but for foreign-made items that contain Chinese rare earths or were produced using Chinese rare-earth technology, with defense uses denied and semiconductor uses reviewed case by case. Several reports and translations of MOFCOM’s notice describe a de minimis trigger at 0.1% of product value, which is intentionally low for complex assemblies and places the burden of proof on exporters; enforcement uncertainty notwithstanding, noncompliance carries the commercial risk of being cut off by Chinese suppliers. This architecture is designed to insert time, documentation risk, and legal uncertainty into multi-tier supply chains, and it explicitly names semiconductor and advanced AI applications as sensitive, raising the probability of selective, time-variable denials.

  Semiconductor equipment is unusually exposed because it embeds rare-earth magnets and related materials in multiple subsystems: high-precision wafer and reticle stages driven by permanent-magnet linear motors; turbomolecular and dry vacuum pumps whose motors and magnetic bearings use rare earths; laser subsystems and magnetron sputtering sources where yttrium, terbium, and gadolinium appear as dopants or target constituents; and numerous ancillary actuators. ASML has already indicated it is preparing for weeks-long shipment delays under the new regime, and Reuters reporting confirms that semiconductor-linked applications will face case-by-case licensing, with blanket denials for defense users. Even short licensing delays propagate as installation and qualification slippage and can push out foundry capacity ramps by 2–8 weeks, particularly for new scanner placements needed to extend 3–5 nm output and to ready 2 nm and High-NA nodes. The core point is not outright scarcity today but the creation of a “shadow quota” via licensing lead-times on parts that lack easy substitution, which raises working-capital needs and the effective cost of time in capex schedules.

  The direct implications for Nvidia operate through 4 channels: wafer supply, packaging and memory availability, China market access, and U.S. tariff exposure on China-origin content. First, Nvidia’s flagship data center GPUs ride TSMC advanced nodes that depend on EUV and deep-UV tool availability; if Chinese-origin magnets or target materials in scanners or deposition tools require new licenses, incremental tool shipments and spares to Taiwan, Korea, the U.S., and Japan can be delayed, slowing incremental wafer capacity adds. Weeks of slippage at ASML translate into later scanner arrival, later installation, and later release-to-production dates at foundries, which in turn shift out Nvidia’s volume ramps on new SKUs at the margin. While today’s supply is already tightly allocated, added frictions extend backlog realization and flatten the near-term slope of unit growth.

  Second, Nvidia’s output is now co-limited by advanced packaging and HBM supply. CoWoS tooling suites and HBM back-end equipment also use precision stages, motors, and vacuum systems with rare-earth content; if license reviews touch these sub-systems or the sputtering targets used in redistribution and passivation steps, packaging capacity adds at TSMC and OSATs can slip. On the memory side, HBM availability is a hard bottleneck for Nvidia module shipments; any licensing drag on tool deliveries or magnet-bearing drive systems for vacuum pumps at SK Hynix, Samsung, or Micron slows HBM ramps. That combination sustains elevated average selling prices but caps near-term shipment upside; revenue becomes more back-half weighted, with gross margins supported by mix and scarcity while opex leverage decouples from unit growth timing. Industry analysis has repeatedly flagged CoWoS and HBM as the binding constraints; a licensing regime that gums up tool flows makes those constraints more persistent.

  Third, the China demand channel is deteriorating on both policy and enforcement. China has intensified customs scrutiny and administrative pressure around Nvidia’s China-specific accelerators, and Beijing can pair rare-earth leverage with procurement guidance steering buyers to domestic alternatives. On the U.S. side, an additional 100% tariff and prospective export controls on “critical software” raise the tail risk that CUDA, drivers, or SDK updates to PRC entities could be restricted, degrading Nvidia’s software monetization and ecosystem lock-in in China. Nvidia’s SEC disclosures show China accounted for a mid-teens share of data center revenue in 2023–2024; while that share has been volatile under shifting controls, the direction of travel under today’s rhetoric is toward lower legal sell-through and higher receivables risk in China. As a result, mix shifts further to U.S. hyperscalers and non-China sovereign demand, supporting aggregate revenue but increasing customer concentration and headline risk.

  Fourth, tariff mechanics on China-origin content affect boards, subassemblies, and server integration. Nvidia has been diversifying final assembly and server integration footprints toward Mexico and other locations, which mitigates U.S. tariff exposure if a 100% levy lands on China-origin hardware, but any residual China content in modules, coolers, fans, or chassis can still encounter substantial import tax or rules-of-origin complexity. The operational response will be accelerated relocation of any remaining China assembly, tighter vendor audits of magnet and target provenance to avoid 0.1% triggers, and higher safety stocks for spares and consumables that risk capture by the new rules. These measures preserve delivery capability but increase working capital and lower operational flexibility, which is a mild headwind to free cash flow conversion while supply remains tight.

  For the broader semiconductor complex, the highest beta sits with equipment suppliers that incorporate high-spec rare-earth magnets and with device makers scaling capacity on aggressive timetables. A few weeks of licensing-driven slippage compounds with tight supplier labor and logistics windows to push revenue recognition out of quarter. Foundries and memory makers will triage tool deliveries toward the most ROI-intensive projects, which arguably helps Nvidia because leading-edge logic for AI accelerators remains top priority; however, that same prioritization can starve less-advanced but still critical back-end tools if their magnet or target content triggers onerous documentation, creating sporadic bottlenecks in substrate, RDL, or probe capacity. Over a 1–2 quarter horizon, this supports a still-tight supply environment for AI accelerators, holding up pricing and margins for Nvidia while capping absolute unit upside and raising the risk that high-volume enterprise adoption is delayed by a few quarters due to lead times on full racks and spares.

  Netting the effects for Nvidia, the median case is a modest negative for near-term unit growth but not an earnings-thesis break: wafer-supply and packaging frictions extend fulfillment timelines by weeks, HBM ramps stay tight and keep ASPs firm, China sell-through shrinks further but is partly offset by ex-China hyperscalers, and higher working capital buffers and compliance costs weigh slightly on cash conversion. The downside case is an escalation where license denials explicitly target semiconductor equipment servicing Nvidia’s key suppliers or where U.S. software controls limit CUDA update distribution to PRC entities; that raises the probability of missed shipments and larger China revenue displacement. The upside case is that MOFCOM grants broad civilian carve-outs with predictable timelines and that U.S. tariffs spare data center components assembled outside China, allowing Nvidia to continue rerouting supply chains; that case largely preserves Nvidia’s delivery outlook while leaving a higher geopolitical risk premium embedded in the multiple. On current information, the skew for Nvidia is to slightly lower volume trajectories with resilient gross margin, elevated backlog durability, and higher operational complexity; for semis as a whole, risk premia rise, dispersion increases, and capex plans encounter higher schedule variance because license queues act as a latent capacity constraint across multiple tool categories.

  Key watch items that will determine whether this becomes a transient compliance tax or a persistent capacity drag are the realized cadence of license approvals for magnet-bearing subassemblies and sputtering targets destined for fabs and OSATs; the degree to which ASML and other toolmakers can certify non-China magnet content or secure waivers; any broadening of U.S. measures to “critical software” that catches CUDA distribution; the response of HBM suppliers to potential licensing friction and their ability to qualify non-China magnet and motor suppliers in vacuum and handling systems; and the extent of China’s administrative guidance against Nvidia hardware in its domestic market. Early indications already point to ASML preparing for weeks of disruption under the re-export clause and to China tightening administrative pressure on Nvidia shipments, which argues for keeping a volatility hedge on Nvidia and on the semis complex until license issuance patterns stabilize

  

​

​

☑️ #271 Nov 3, 2025

Tubi CEO on Streaming Landscape, Google’s AI Strategy, Starcloud’s Space GPUs | Nov 3, 2025

@theinformation4080: [6:30] We also talk with StarCloud Co-Founder & CEO Philip Johnston about launching an NVIDIA H100 GPU into space, its core cooling technology, and their business model.

+ Related content:

• blogs.nvidia.com (10/15/25): [Excerpt] How Starcloud Is Bringing Data Centers to Outer Space.

  The NVIDIA Inception startup projects that space-based data centers will offer 10x lower energy costs and reduce the need for energy consumption on Earth.

  Starcloud’s upcoming satellite launch, planned for November, will mark the NVIDIA H100 GPU’s cosmic debut — and the first time a state-of-the-art, data center-class GPU is in outer space.

  The 60-kilogram Starcloud-1 satellite, about the size of a small fridge, is expected to offer 100x more powerful GPU compute than any previous space-based operation.

  • @Starcloud_Inc1 (11/3/25): Thanks for the ride @SpaceX! Beautiful shot of Starcloud-1 deploying there!

    @PhilipJohnst0n, @ezrafeilden, @adi__oltea

    • @SpaceX (11/2/25): Deployment of all payloads confirmed.

      

      

☑️ #270 Nov 3, 2025

An updated overview of the situation regarding Nvidia Corporation’s semiconductor exports to China

chat.openai.com (chatgpt) > [Sources not verified. Do your own research]:

The U.S. has imposed export controls on advanced AI and high-performance computing chips to China. For example, in October 2023 the U.S. Department of Commerce announced restrictions on the export of Nvidia’s A800 and H800 chips (among others) to China. (CNBC)

Nvidia, in an April 2025 filing, said these controls would cost them an estimated US$5.5 billion in charges, especially because of the H20 chip inventory and previously covered sales now subject to restrictions. (euronews)

Nvidia’s CEO, Jensen Huang, has said that U.S. export controls to China have been “a failure” because China is making rapid progress. (DW News)

Recent developments & nuance

• The export control regime is not simply a total ban: many chips remain export-eligible if they meet certain performance thresholds and are appropriately licensed. For instance, regulating performance density, total processing performance (TPP), etc. (CSIS)

• The U.S. appears to be shifting somewhat: The licence-granting process for the H20 chip has begun, allowing Nvidia to hope for resumed exports to China. (CNBC)

• On the Chinese side, there are reports of tighter customs scrutiny of chip imports, particularly of U.S.-made AI processors. (This reflects a broader push by China toward self-sufficiency in semiconductor/AI technology.) (Financial Times)

Nvidia can export certain chips to China — particularly ones designed to comply with U.S. export rules (e.g., the H20) — but many of their highest-end AI chips (e.g., H100, H800, etc) are under tight restrictions.

+ Related content:

• ft.com (update; 11/3/25): China offers tech giants cheap power to boost domestic AI chips.

  Local governments in Gansu, Guizhou and Inner Mongolia offered to cover up to 50% of electricity costs for data centers, provided they use domestically produced AI processors, such as those developed by Huawei and Cambricon.

  China is offering substantial power subsidies to major tech firms like:

  • ByteDance

  • Alibaba (9988.HK)

  • Tencent (0700.HK)

☑️ #269 Oct 31, 2025

Samsung Teams With NVIDIA To Lead the Transformation of Global Intelligent Manufacturing Through New AI Megafactory

news.samsung.com: [Excerpts] Korea on October 31, 2025. Driving company-wide digital transformation through AI-powered manufacturing across semiconductors, mobile devices and robotics.
Leveraging over 50,000 NVIDIA GPUs and the NVIDIA Omniverse platform to scale next-generation AI manufacturing infrastructure.
Advancing manufacturing and humanoid robotics with NVIDIA AI platforms to enable greater intelligence and autonomy.

From 25+ Year Collaboration to a Stronger AI Chip Alliance

Samsung and NVIDIA celebrate a collaboration that spans more than 25 years, beginning with Samsung’s DRAM powering NVIDIA’s early graphics cards and extending through foundry partnership.

In addition to their ongoing collaborations, Samsung and NVIDIA are also working together on HBM4. With incredibly high bandwidth and energy efficiency, Samsung’s advanced HBM solutions are expected to help accelerate the development of future AI applications and form a critical foundation for manufacturing infrastructure driven by these technologies.

Built with the company’s 6th-generation 10-nanometer (nm)-class DRAM and a 4nm logic base die, Samsung HBM4’s processing speeds can reach 11 gigabits-per-second (Gbps), far exceeding the JEDEC standard of 8Gbps.

Samsung will also continue to deliver next-generation memory solutions, including HBM, GDDR, and SOCAMM, as well as foundry services, driving innovation and scalability across the global AI value chain.

+ Continue reading

☑️ #268 Oct 29, 2025

Exclusive: Nvidia backs new data center to reduce electricity spikes

axios.com: [Excerpt] A new kind of data center built by a coalition including Nvidia aims to smooth out power use as AI demand surges.

Why it matters: Shared exclusively with Axios, the project is the first commercial rollout of software that adjusts energy draw in real time — a model the companies say can ease strain on the grid and curb electric costs.

Nvidia will deploy new software developed by Emerald Al to reduce power use at the Aurora data center in Virginia.  

+ Continue reading

+ Related content:

• linkedin.com > Emerald AI (): A new era begins for AI and energy. 🚨
  
  🚀 Today, Emerald AI, NVIDIA, EPRI, Digital Realty, and PJM Interconnection announced the world’s first power-flexible AI Factory at commercial scale: the 96 MW Aurora AI Factory in Virginia.
  
  This milestone project establishes the reference architecture for how AI infrastructure and electric grids can work together to supercharge innovation while maintaining reliability.
  
  ⚡ Power-flexible AI Factories have the potential to unlock 100 GW of existing grid capacity to fuel the AI revolution without the need for massive new power plants. This approach keeps electricity affordable, strengthens grid reliability, and supports sustainable growth.
  
  Virginia Governor Glenn Youngkin recognized Aurora as a model for “promoting economic development and AI competitiveness while at the same time protecting the affordability and reliability of the power grid.”
  
  🌍 So the message is clear: AI and the electricity grid are no longer on separate tracks. Together, we’re defining the blueprint for power-flexible AI infrastructure. The AI revolution is the next moonshot, and power-flexible AI Factories can unlock energy abundance for AI.
  
  🗞️ Read Amy Harder’s exclusive in Axios: https://lnkd.in/ehBMs2aT
  
  🔗 Read the full press release here: https://lnkd.in/eXgi-6rg
  
  A huge heartfelt thank you to our phenomenal collaborators—NVIDIA, Digital Realty, EPRI, PJM Interconnection, and all our partners and supporters!
  hashtag#AI hashtag#GridFlexibility hashtag#EmeraldAI hashtag#AIRevolution hashtag#DataCenters

  ​

• companiesmarketcap.com Nvidia became the first company to reach a $5T market valuation on Wednesday (10/29)

  

☑️ #267 Oct 28, 2025 🟠 opinion

AI is about Reindustrialization

@ssankar: Jensen and @nvidia delivering the AI infrastructure for U.S. manufacturing dominance. Accelerated compute turbocharging AIP—from the Lowe’s supply chain to the frontline.

AI isn’t about AI – it is our asymmetric advantage to deliver country-scale transformation. AI is about Reindustrialization. AI is how we give the American worker superpowers and make them 50x more productive.

American leadership in industrialization is re-emerging: Mobilize.

• @ssankar (10/9/25): 1/ We are in an undeclared state of emergency. America’s adversaries are circling, and the American industrial base that dominated the 20th century is dormant. For the CCP, it is not enough for China to prosper... America must fall. It’s time to Mobilize. Coming March 2026.

  

⚡️

@nvidianewsroom: #NVIDIAGTC: Palantir and NVIDIA team up to operationalize AI — turning enterprise data into dynamic decision intelligence. NVIDIA accelerated computing, CUDA-X libraries, open-source Nemotron models are now available via the Palantir Ontology and AIP.

https://nvda.ws/4nnRqZL

☑️ #266 Oct 16, 2025

Largest Company in the world by market‑capitalization at market close on December 31, 2025?

chat.openai.com (chatgpt) > [Sources not verified. Do your own research]:

✅ Why the chance is reasonably high

Momentum matters: If NVIDIA retains its lead into late 2025 (assuming no major disruption), then the odds of still being #1 at year‑end are solid.

• Competition and headwinds: Even if NVIDIA is ahead now, other large firms like Microsoft Corporation, Apple Inc., Alphabet Inc. etc. have enormous scale and could pull ahead with a strong quarter, favourable regulation, or a major innovation. Also risk factors like export restrictions (for AI chips) could weigh.

🧭 What to Watch (Rest of 2025)

1. Nvidia Earnings (Q4 2025 due ~December)

   • A blowout or disappointing earnings report in late 2025 could shift the market cap ranking right before December 31.

   • Keep tabs on revenue growth in AI/data center, and any guidance changes.

2. Microsoft / Apple Moves

   • Any product surprise, strategic shift, or strong quarter from them could tighten the race. Both are within striking distance.

3. Macro Events

   • Rate cuts, geopolitical tensions, or changes in chip regulations (e.g., U.S.-China policy) could disproportionately affect NVIDIA.

4. Year-End Trading Trends

   • Tax-loss harvesting, window dressing by funds, or rebalancing can create volatility in December.

+ Related content:

• @hhuang: The market will soon recognize that this is the only TPU that can compete with NVIDIA.

  8 years ago, I attended a tech talk by Jeff Dean at Stanford about TPUv2, which was impressive as they started to use it for ML training at the time. So the gap between Trainium 3 and TPUv7 is much larger than people expected, and this gap may widen due to AMZn’s “cost-first, speed-second” approach.

  I bet long Goog short Amzn will be very popular after this ER season.

  • @zephyr_z9 Google officially starts selling TPUs to external customers and competes directly with Nvidia now

    • @zephyr_z9: Broadcom’s 5th customer ($10B) isn’t Apple or XAI It’s Anthropic. They won’t design a new chip. They will be buying TPUs from Broadcom. Expect Anthropic to announce a funding round from Google soon

      

☑️ #265 Oct 16, 2025

$20B Raise via SPV: Bold, but Possible

theinformation.com: [Excerpt] Rather than buying all the NVIDIA’s GPUs outright, xAI Corporation is leaning on an SPV leasing model.

xAI’s unusual dealmaking to fund Musk’s Colossus 2. xAI Corporation plans to raise $20 billion for a deal using a special purpose vehicle (SPV) that would acquire semiconductors from Nvidia Corporation and then lease them to xAl, which will have the option to buy them at the end of the lease.

The SPV will reportedly comprise $7.5 billion in equity, anchored by Valor Equity Partners, and $12.5 billion in debt financing.

+ Continue reading

+ Related content:

• @markets: Nvidia to Invest in Musk’s xAI.

  Elon Musk’s artificial intelligence startup xAI is raising more financing than initially planned, tapping backers including Nvidia to lift its ongoing funding round to $20 billion, according to people familiar with the matter. Nvidia is investing as much as $2 billion in the equity portion of the transaction, the people said. Bloomberg’s Annabelle Droulers reports.

• perplexity.ai > [Sources not verified. Do your own research]: Vertical Integration Without Full Ownership

  This separates asset ownership (capital-intensive) from operational use (compute at scale), keeping xAI more agile and less burdened by balance-sheet liabilities. It’s a clever workaround in a tight capital market and reflects a financial-engineering-first approach, led by people like Antonio Gracias—who has deep roots in Tesla’s early scaling.

  Scaling Compute as a Moat

  Going from 0 to 200,000 GPUs in just over 200 days is practically unheard of. That puts xAI’s Colossus ahead of all but a few hyperscalers in terms of raw compute, making this infrastructure the central competitive moat for LLM training and inference at scale.

  Whoever owns (or controls) the most compute wins.

  In the AI arms race, models matter—but scale, speed of iteration, and compute density are increasingly the differentiators. xAI clearly sees this and is making infrastructure its front line.

  Nvidia Dependency: Strength or Risk?

  Nvidia wins huge in this deal—again. But it also underscores the systemic dependency emerging in AI ecosystems on Nvidia’s hardware stack (H100s, GH200s, and possibly Blackwell next). For xAI, this means:

  • Performance edge if they maintain early access to Nvidia’s bleeding-edge.

  • Risk if Nvidia’s supply chain gets pinched, prices rise, or competitors (e.g., AMD, Intel, or open ASIC projects) catch up.

  This dependency could eventually become a strategic liability unless xAI hedges it long term (e.g., by co-developing custom silicon, as Google and Amazon have done).

  • securitiex.substack.com: [Excerpt] Scaling Compute as a Moat:

    Going from 0 to 200,000 GPUs in just over 200 days was practically unheard of. That puts xAI’s Colossus ahead of all but a few hyperscalers in terms of raw compute, making this infrastructure the central competitive moat for LLM training and inference at scale.

    The core idea: Whoever owns (or controls) the most compute wins.  
    
    In only 122 days, xAI built its first huge data center, Colossus in Memphis, Tennessee, equipping it with 100,000 Nvidia GPUs. 92 days later, capacity was boosted to 200,000 units. Now, Antonio Gracias (Valor Equity Partners) anchors xAI Corporation’s plan to raise $20 billion for a deal using a special purpose vehicle (SPV) that would acquire semiconductors from Nvidia Corporation and then lease them to  Colossus 2. 

☑️ #264 Oct 16, 2025

Data Center Credits and 800 VDC Power Architecture

chat.openai.com (chatgpt) > [Sources not verified. Do your own research]:

Below is a deep‑dive sketch of how one can view NVIDIA’s modern ecosystem as a flywheel / virtuous loop — how each component (CoreWeave, power architecture, AI factories, GPU cluster partners, system alignment) not only creates value, but amplifies and compounds it. It’s not a perfect public blueprint, but a synthesis of observable moves, technology signals, and leverage points.

NVIDIA + Data Center Credits: co-risk

• CoreWeave as a GPU Cloud Partner / Anchor Tenant
  NVIDIA has strategically invested in and backed CoreWeave, including via equity, guaranteed capacity purchases, and preferential access to GPU hardware. For example, NVIDIA committed to purchase unsold compute capacity from CoreWeave under a $6.3B agreement through 2032, which de‑risks CoreWeave’s utilization and gives NVIDIA a backstop on wastage. (Reuters)
  This gives NVIDIA not only a downstream outlet for its GPUs, but also control/visibility into a high‑volume deployment partner.

• Data Center Credits / Financial Guarantees
  By providing “credits” or guaranteed capacity purchase, NVIDIA effectively subsidizes the risk of building large GPU farms. That lowers the hurdle for scaling infrastructure, encouraging more aggressive buildouts by CoreWeave.
  It aligns incentives: CoreWeave pushes GPU adoption which in turn drives more orders and more GPU volume for NVIDIA.

  • Because CoreWeave has a backstop, it can carry more speculative capacity, absorb utilization dips, or invest more aggressively in new data centers, confident that NVIDIA will anchor some of the demand.

  • By embedding financial instruments (like data center credits, guaranteed capacity purchases, credit lines) into partnership agreements, NVIDIA aligns incentives, de-risks investment for partners, and ensures partners are motivated to scale fast (because the downside is shared).

    • These instruments also help NVIDIA “steer” capacity (e.g. ensuring demand flow, smoothing utilization, encouraging certain geographies).

    • CoreWeave can aggressively build GPU farms (less risk, more capital confidence).

AI Factories & 800 VDC Power Architecture: viable farms

This is the “infrastructure engine” of the ecosystem: making ultra‑dense, efficient AI compute farms viable.

Throughout, the 800 VDC power architecture is a critical enabler — without the power scaling, the density and efficiency gains needed for hyper-scale AI factories would falter. Similarly, data center credits / financial risk‑sharing ensure that partner capacity expansions can move faster than they would under purely market risk. And system alignment & orchestration make sure the ecosystem does not fragment, reducing friction and reinforcing compounding effects.

• NVIDIA is promoting an 800 VDC power distribution architecture for AI factories (data center farms) to replace less efficient AC or multi-stage conversion DC systems. (NVIDIA Developer)

• The shift to 800 VDC allows fewer conversions (less loss), higher power density per conductor, reduced copper usage, fewer components, lower cooling overhead, and improved reliability. (NVIDIA Developer)

• At scale (megawatt+ racks), traditional 54 V or AC power architectures hit physical and thermal limits; 800 VDC is more scalable and future-proof. (NVIDIA Developer)

• NVIDIA envisions these as “AI factories,” where power, cooling, compute, orchestration are co‑engineered holistically.

  • The superior power architecture reduces TCO and allows NVIDIA’s GPU systems to be placed in much denser, more efficient configurations. That makes NVIDIA’s proposition more compelling vs. alternatives (other chips, custom accelerators, etc.).

  • This enables partners and cloud providers to adopt NVIDIA’s higher density racks, knowing the underlying power system is optimized.

  • The increased efficiency and capacity reduce the cost per FLOP (or effective compute), which further accelerates demand — which in turn prompts NVIDIA to push further innovations (e.g. better power management, cooling, fault tolerance).

  • As power systems evolve, NVIDIA can capture value in the design, components, and standards (e.g. power converters, DC distribution modules) — effectively extending its value chain upward into the power stack.

    • Because NVIDIA can plan ahead, it positions the system for high density and efficiency.

    • The kind of energy demand at gigawatt scale will stress grid infrastructure, regulatory constraints, and energy cost volatility. The 800 VDC + energy storage elements must be robust.

Thus, the AI factory + 800 VDC power architecture is the infrastructure backbone that unlocks the densification and scale required to make the rest of the flywheel feasible and cost-effective.

Deploys / enables AI factories (internal or via partners) built around those systems

NVIDIA doesn’t simply sell a GPU and walk away; rather, they often co-engineer with partners (e.g. in rack / system integration, power delivery, cooling). This ensures that the hardware and the system are well tuned.
For example, when designing the 800 VDC ecosystem, NVIDIA works with power electronics, rack vendors, component suppliers. (NVIDIA Developer)

• Partners or invests in GPU cloud/infrastructure players (CoreWeave, Lambda, Nebius, Oracle, etc.) who adopt NVIDIA’s tech and buy scale.

  These GPU cluster providers drive utilization scale — the more deployed GPU infrastructure, the more compute-hour demand is generated, reinforcing the value of NVIDIA’s hardware and software.

  They provide market feedback signals (evenues, usage, feedback, and demand signals) — which architectures are working, where bottlenecks lie, what new features (memory, interconnect, reliability) customers demand; NVIDIA can feed that learning back into its R&D.

  • Lambda
    Lambda offers “AI Factories” as a service: single-tenant GPU clusters, co-engineered deployments, scaled to tens of thousands of GPUs, in multiple data centers. (lambda.ai)
    They adopt NVIDIA’s newer GPU systems and presumably will embrace the advanced power and rack architectures.

  • Nebius
    Reportedly, Microsoft signed a large contract with Nebius to lease AI compute power (tied to NVIDIA’s chips), making Nebius one of the “neocloud” GPU infrastructure providers. (Financial Times)
    Nebius is part of the emerging GPU cloud ecosystem, benefiting from scale, specialization, and integration with NVIDIA’s roadmaps.

  • Oracle + NVIDIA collaboration
    Oracle Cloud Infrastructure (OCI) and NVIDIA are integrating their stacks: 160+ NVIDIA AI tools and microservices will be available natively in OCI. (NVIDIA Newsroom)
    Oracle becomes a cloud partner/reseller/distributor of NVIDIA’s AI stack, increasing reach into enterprise workloads.

• Software stack alignment and platform bundling
  By embedding its AI software (CUDA, libraries, inference stacks, microservices) into partner clouds (OCI, CoreWeave, etc.), NVIDIA ensures that the compute hardware is tightly integrated with the software environment, lowering friction and increasing lock-in. The more workloads run on NVIDIA’s stack, the harder it is to migrate away.

• Interoperability & standardization
  These cluster providers adopt or co-develop around NVIDIA’s architectures (rack architecture, power, network fabrics). For example, NVIDIA is pushing the 800 VDC design and referencing that partners like CoreWeave, Lambda, Nebius, Oracle are also moving in that direction. (eWeek)
  Oracle/OCI embedding of NVIDIA tools means that when enterprises use Oracle, they are effectively consuming NVIDIA’s software stack, which increases lock-in and synergy.

  • Open standards, reference architectures, and ecosystem contributions

    • NVIDIA is pushing MGX, rack reference architectures, power system designs, and contributions into open consortia (e.g. via OCP) so that partners can more easily align. (MEXC)
      By making these open or semi‑standard, adoption friction lowers, and ecosystem actors can plug in more easily.

    • NVIDIA must balance protecting its IP and capturing margins with promoting enough openness so that partners adopt its architectures. If the ecosystem sees it as too closed, adoption friction may rise.

☑️ #263 Oct 15, 2025

AI INFRASTRUCTURE PARTNERSHIP (AIP), MGX, AND BLACKROCK’S GLOBAL INFRASTRUCTURE PARTNERS (GIP) TO ACQUIRE ALL EQUITY IN ALIGNED DATA CENTERS

aligneddc.com: [Excerpt] AIP was founded by BlackRock, Global Infrastructure Partners (GIP), a part of BlackRock, MGX, Microsoft, and NVIDIA to expand capacity of AI infrastructure and help shape the future of AI-driven economic growth. Its financial anchor investors include the Kuwait Investment Authority and Temasek.

In less than a decade, Aligned has evolved into one of the largest and fastest growing data center companies globally. The Company designs, builds, and operates cutting-edge data campuses and data centers for the world’s premier hyperscalers, neocloud, and enterprise innovators. Aligned’s portfolio includes 50 campuses and more than 5 gigawatts of operational and planned capacity, including assets under development, primarily located in key Tier I digital gateway regions across the U.S. and Latin America including Northern Virginia, Chicago, Dallas, Ohio, Phoenix, Salt Lake City, Sao Paulo (Brazil), Queretaro (Mexico), and Santiago (Chile).

+ Continue reading

+ Related content:

• chat.openai.com (chatgpt): [Sources not verified. Do your own research]:

  Key Strategies Nvidia is Using to Scale Data Centers / AI Infrastructure

  1. Reference Architectures & Modular Systems (MGX, HGX, GB200, NVL72 etc.)
     Nvidia is pushing standardized, modular building blocks so that customers and partners can more cheaply and more quickly build large AI‐/HPC‐capable data centers.

     • Their Blackwell platform (GPUs, CPUs, networking) is designed with openness in mind. They’ve contributed rack / thermal / cooling / networking mechanical/electrical design to the Open Compute Project (OCP). (investor.nvidia.com)

     • The MGX modular reference architecture supports many server variations (size, cooling, compute etc.) to suit different scale needs. (NVIDIA Newsroom)

       • Modular architecture (MGX), liquid cooling, high compute density per rack, giving other vendors / operators access to NVIDIA’s design helps scale more uniformly, reduces reinventing, increases possibility of optimized data center builds. Useful in adaptive scaling. (NVIDIA Investor Relations)

     • The GB200 NVL2 and NVL72 systems provide very dense compute + memory + networking, aimed at large LLM inference, retrieval tasks, etc. These help hyperscalers build more capable AI racks/campuses. (investor.nvidia.com)

       • NVIDIA contributed design specs (rack architecture, thermal / electro‑mechanical design, etc.) for its GB200 NVL72 (which is a very high density compute platform) to the Open Compute Project (OCP). (NVIDIA Investor Relations)Performance claims (for the platform) include e.g. 30× faster real‑time trillion‑parameter inference vs past generation GPU; but metrics on facility level (PUE, WUE) not specified in the public contribution. (NVIDIA Investor Relations)

  2. Power, Cooling, and Efficiency Innovations

     Scaling isn’t only about GPUs. It’s also about how to deliver power, deal with heat, reduce losses, keep energy and operational costs manageable. Nvidia is doing several things here:

     • Pushing 800 VDC (High Voltage Direct Current) power systems, to enable 1‑megawatt+ server racks. This reduces losses, simplifies power conversion, lowers copper mass, etc. (NVIDIA Developer)

       • Moving from traditional AC/DC power distribution to high‑voltage DC (800 VDC) to reduce losses and enable more efficient power delivery into racks. The roll‑out of the 800 VDC architecture starting around 2027 to support 1 MW+ racks.

         Reduced energy losses via fewer AC/DC conversions; less copper, less thermal loss in power delivery; better scalability for very high power racks; likely reduced cooling overhead (because less waste heat from inefficient conversions) and lower maintenance. (NVIDIA Developer)

     • Collaborations with power components firms (Infineon, STMicroelectronics, Texas Instruments, etc.) and data center power system operators (Eaton, Vertiv, Schneider, etc.) to build the supporting supply chain / gear. (DataCenterDynamics)

       • Schneider Electric + NVIDIA: AI Factories at Scale (Europe R&D / Rack / Cooling / Power / Controls)

         Partnership announced in mid‑2025: developing rack systems, cooling, power, BMS (building management/control systems) for “AI factories” (i.e. large‑scale data centers) across Europe and beyond. Focus on high‑density racks; improved power & cooling designs; more efficient controls; potentially more energy‑efficient cooling/power infrastructure; alignment with EU policies (“AI Continent Action Plan”, InvestAI, etc.). Adaptive design that can scale. (Schneider Electric)

         Not many public concrete numbers (e.g. PUE, energy savings) released yet; likely still in R&D / early development.

     • Datacenter Energy Optimized Power Profiles (Blackwell B200 Feature)

       A recent NVIDIA software / control‑feature (on Blackwell architecture) to allow data centers to choose power profiles to manage energy vs performance trade‑offs. (arXiv)Enables respecting facility power constraints, optimizing for energy usage: reducing energy while maintaining performance; helps in high density / constrained environment. This kind of adaptive scaling (in workload / power profile) supports operational efficiency, fewer overprovisioned resources. (arXiv)In their reported tests: up to 15% energy savings while maintaining >97% of performance, and up to 13% throughput increase in power‑constrained facilities. That is significant. (arXiv)

     • Using liquid cooling, advanced thermal & rack design, and densification with high bandwidth networking to keep rack footprint vs performance efficient. Also sharing design specifications via OCP to help partners build consistent infrastructure. (investor.nvidia.com)

       • nVent & NVIDIA: AI‑Ready Liquid Cooling Solutions

         Collaboration between NVIDIA and nVent (liquid cooling hardware vendor) to deploy liquid cooling at scale, supporting NVIDIA’s GB200 / NVL72 and next‑gen platforms. (datacenterfrontier.com)

         Designed reference architectures for high density racks (NVL36 / NVL72), with custom cooling manifold, coolant distribution units (CDUs), and liquid‑to‑air exchangers. Allows better thermal management, higher compute density, reduced cooling overhead.

  3. Partnerships & Ecosystem Expansion

     Nvidia is not trying to build all data centers themselves. They’re enabling or partnering with:

     • Hyperscalers, cloud providers, enterprises who need GPU/AI infrastructure. For example, their HGX partner program (ODMs like Supermicro, GIGABYTE, etc.) to build AI systems. (ir.tdsynnex.com)

     • Collaborations with traditional telcos or network operators, for distributed AI data centers (SoftBank in Japan, etc.) where AI/5G/edge workloads are combined. (NVIDIA Newsroom)

     • Certified AI‑ready data centers & colocation partners. For example in India: Sify’s campuses certified under NVIDIA’s “DGX-Ready Data Center” program, offering pay‐per‐use for AI workloads. (techfocusasia.com)

  4. Networking & Interconnect Technology

     As data centers / AI factories scale, the networking becomes a major axis of differentiation.

     • Spectrum‑X Ethernet platform: high throughput, adaptive routing, congestion control, etc., is being deployed. (investor.nvidia.com)

     • NVLink interconnects and designs (e.g., NVLink cable modules) are being shared via OCP for better compatibility and densification. (investor.nvidia.com)

  5. “AI Factory” / Data Center as a Product: Better Tools & Platforms

     Nvidia frames large data centers as “AI factories”—not just places to run compute, but integrated systems. This includes:

     • Tooling for storage + data infrastructure that are AI‑aware. E.g. the NVIDIA AI Data Platform adds accelerated computing (GPUs, DPUs) to storage systems, optimizing for AI query access. (investor.nvidia.com)

     • Enhanced enterprise server options so organizations can start with smaller clusters or “AI ready” nodes and scale up without redesigning everything. (NVIDIA Newsroom)

  6. Large Investments & Strategic Deals

     • Nvidia is investing in, or backing entities that build large-scale AI infrastructure, or making long-term supply agreements. (Some of the “news” items show big purchases/commitments). (Reuters)

     • They are aligning with energy, financing, power supply, and real estate needs (the non‑GPU parts) via large deals and partnerships. The $40B acquisition of Aligned Data Centers via a consortium including Nvidia is one example. (Reuters)

☑️ #262 Oct 14, 2025

Early access: NVIDIA DGX Spark

@lmsysorg: SGLang In-Depth Review of the NVIDIA DGX Spark is LIVE!

Thanks to@NVIDIA’s early access program, SGLang makes its first ever appearance in a consumer product, the brand-new DGX Spark.

The DGX Spark’s 128GB Unified Memory and Blackwell architecture set a new standard for local AI prototyping and edge computing. We’re thrilled to bring these cutting-edge performance insights and software support to the developer community.

Our review dives into how to efficiently deploy and accelerate large models like Llama 3.1 70B, GPT-OSS using SGLang’s EAGLE3 speculative decoding and @Ollama on this beautiful piece of engineering.

Unboxing video and tech blog in the thread #SGLang #NVIDIA #SparkSomethingBig #Blackwell #DGXSpark #AIInference #LLMServing

@lmsysorg: Special thanks to @nvidia @NVIDIAAI @NVIDIAAIDev for the early access program and community member @yvbbrjdr @richardczl for the review!

+ Related content:

• lmsys.org(10/13/25): [Excerpt] Around the back, the DGX Spark offers an impressive array of connectivity options: a power button, four USB-C ports (with the leftmost supporting up to 240 W of power delivery), an HDMI port, a 10 GbE RJ-45 Ethernet port, and two QSFP ports driven by NVIDIA ConnectX-7 NIC capable of up to 200 Gbps. These interfaces allow two DGX Spark units to be connected together, allowing them to run even larger AI models.

  The use of USB Type-C for power delivery is a particularly interesting design choice, one that’s virtually unheard of on other desktop machines. Comparable systems like the Mac Mini or Mac Studio rely on the standard C5/C7 power connector, which is far more secure but also bulkier. NVIDIA likely opted for USB-C to keep the power supply external, freeing up valuable internal space for the cooling system. The trade-off, however, is that you’ll want to be extra careful not to accidentally tug the cable loose

  

  • @petergostev: Nvidia’s DGX Spark goes on sale today and @lmsysorg

    have done a brilliant bit of benchmarking vs other systems.

    In short, it is a very usable system for smaller models and closer in performance to Apple’s devices (e.g. Mac Mini M4 Pro), but it is priced at $4,000 vs $1,400 for Mac Mini M4 Pro. It is also not as powerful as more dedicated GPUs like RTX 5090.

    Worth reading & watching the full review in the original post.

    

☑️ #261 Oct 14, 2025

NVIDIA’s major recent developments

chat.openai.com (chatgpt) and phind.com (Phind-70B) > [Sources not verified. Do your own research]:

Major Recent Developments

• DGX Spark “personal AI supercomputer” launch: NVIDIA has begun selling a compact, desktop‑sized AI system (DGX Spark) that packs high-end AI performance (for models up to ~200 billion parameters) for individual or small‑scale use. (The Verge)

  • DGX Spark & “Station”: The miniaturized AI supercomputer (DGX Spark) goes on sale Oct 15, 2025; a higher-end “Station” is teased.

    The world’s smallest AI supercomputer capable of petaflop performance.    

  • Edge, agentic AI, and personalized AI models: The DGX Spark and smaller AI appliances point toward distributed, on-device, and hybrid AI compute markets.

  • @nvidianewsroom (10/14/25): NVIDIA and its partners will start shipping DGX Spark — the world’s smallest AI supercomputer. Early recipients are testing, validating and optimizing their tools, software and models for DGX Spark. Built on the NVIDIA Grace Blackwell architecture, DGX Spark integrates NVIDIA GPUs, CPUs, networking, CUDA libraries and NVIDIA AI software, accelerating agentic and physical AI development.

    

    • @nvidianewsroom (10/13/25): To celebrate DGX Spark shipping worldwide starting Wednesday, our CEO Jensen Huang just hand-delivered a DGX Spark to @elonmusk, Chief Engineer at @SpaceX, today in Starbase, Texas. The exchange was a connection to the new desktop AI supercomputer’s origins — the NVIDIA DGX-1 supercomputer — as Musk was among the team that received the first DGX-1 from Huang in 2016. See our live blog for details, and on more early deliveries to developers, creators, researchers, universities, and more: https://blogs.nvidia.com/blog/live-dgx-spark-delivery

      

      

• Power supply / energy efficiency push: NVIDIA is pushing data center power distribution toward 800‑volt direct current (DC) schemes, with partner Power Integrations joining to deliver GaN power chips to reduce losses. (Reuters)

• US export license for UAE shipments: The U.S. government granted NVIDIA permission to export advanced AI GPUs (e.g. Blackwell family) to the United Arab Emirates, opening a significant new regional channel. (Tom’s Hardware)

• China customs crackdown: Chinese authorities have increased inspections of AI chip imports, including NVIDIA products, as part of a broader move to limit dependence on U.S. semiconductor supply. (Financial Times)

• SoftBank / Japan collaboration: NVIDIA is working with SoftBank to build AI + 5G telecom infrastructure in Japan, including edge AI services. (NVIDIA Newsroom)

• Saudi PIF / HUMAIN AI factories: A strategic partnership to build large AI “factories” in Saudi Arabia under the HUMAIN initiative, deploying hundreds of thousands of GPUs over coming years. (NVIDIA Newsroom)

• Investment in Intel: NVIDIA took a 5 % stake in Intel (~$5 billion), positioning itself as a major external shareholder and enabling tighter collaboration (e.g. NVLink for x86) between the two firms. (Network World)

• Talent / acquisition moves:

  • NVIDIA licensed or acquired interconnect technology from Enfabrica and reportedly hired key personnel in that domain. (Network World)

  • NVIDIA continued a string of software / AI startups acquisitions in 2025 (e.g. Gretel, Lepton AI, CentML). (CRN)

• Silicon photonics / networking: NVIDIA introduced silicon photonic network switches and co‑packaged optics (CPO), integrating optics into switch hardware to improve performance and efficiency. (Network World)

• Software & AI tooling updates:

  • General availability of NeMo microservices for modular AI model building. (Network World)

  • Launch of AgentIQ toolkit for connecting AI agents and orchestration. (Network World)

  • Announcement of Dynamo (open‑source inference runtime) for efficiency in AI model deployment. (Network World)

• Energy optimization at GPU / data center level: New features in Blackwell chips allow workload‑aware power profiles, potentially yielding ~15 % energy savings while preserving ~97 % performance for critical tasks. (arXiv)

• Expansion of GeForce NOW gaming platform with significant content additions.

Strategic Partnerships

• OpenAI: NVIDIA and OpenAI are collaborating to deploy at least 10 gigawatts of NVIDIA systems for next-gen AI infrastructure, with first phases expected in H2 2026. (Network World)

• Telecom / sovereign AI: In Europe, NVIDIA is working with Orange, Telenor, Telefónica, Swisscom, Fastweb to build sovereign AI infrastructure in multiple countries. (investor.nvidia.com)

• Saudi PIF / HUMAIN: As above, leveraging national-scale AI infrastructure investment. (NVIDIA Newsroom)

• SoftBank / Japan: Integrating AI + 5G telco infrastructure, building AI marketplace, edge services. (NVIDIA Newsroom)

• Intel: As a 5 % shareholder, NVIDIA is working with Intel to integrate NVLink into x86 CPUs and co‑develop custom data center and PC products. (Network World)

• Fujitsu: Collaboration on vertical industry AI agent platforms (e.g. healthcare, manufacturing, robotics). (Network World)

• Cisco: The partnership is expanding to tie Cisco’s Silicon One tech with NVIDIA’s Ethernet / networking stack for enterprise AI deployment. (fxtechnology.org)

• IBM / Oracle / cloud providers: NVIDIA is integrating its AI stack into hybrid cloud — e.g. Watsonx with NVIDIA, plus making its AI Enterprise platform native in Oracle Cloud Infrastructure. (Network World)

  • Network infrastructure deals with Meta and Oracle for Spectrum-X Ethernet switches

• Partnership with Microsoft Azure for world’s first GB300 NVL72 supercomputing cluster.

  • Blackwell Ultra / GB300 / NVL72: The new generation (Ultra / GB300) is tailored for large language models, reasoning, inference workloads, and dense AI compute. The NVL72 rack architecture (72 GPUs + 36 Grace CPUs) offers extreme compute density. (CRN)

    NVIDIA Blackwell is setting new benchmarks in AI inference. 

Other Product Innovations

• Energy efficiency improvements of 100,000x in LLM inference over past decade.

• Silicon photonic / co‑packaged optics: To reduce latency, power, and improve efficiency in AI data centers. (Network World)

• Networking / interconnect (NVLink, NVLink Fusion, Spectrum‑XGS, Spectrum‑X): High‑bandwidth interconnects between GPUs, CPUs, and across nodes are central to NVIDIA’s scaling vision. (Network World)

• NeMo microservices, AgentIQ, Dynamo: Software stacks and toolkits for building, serving, and orchestrating AI models and agents. (Network World)

• Energy / power management: On‑chip power profiles and data center energy optimizations built into architecture (e.g. Blackwell) to improve efficiency. (arXiv)

• Synthetic / data privacy tools: Acquisition of Gretel (synthetic data), CentML (model optimization), etc., to bolster data pipelines and AI model performance. (CRN)

• Quantum / research: NVIDIA is establishing an Accelerated Quantum Research Center, signaling interest in qubit / hybrid quantum‑classical architectures. (Network World)

More Global Expansion

• Europe / AI sovereignty: Strategic initiatives across Europe for tenfold increase in AI computing power. It is placing AI centers in Germany, Italy, Spain, U.K., Finland, Sweden to support research, training, infrastructure, and sovereign AI deployment. (investor.nvidia.com)

  • £2 billion investment commitment to UK AI startup ecosystem.

  • Partnerships with European telcos (Orange, Telefónica, Telenor, etc.) to roll out regional AI fabrics and edge compute. (investor.nvidia.com)

• Middle East / UAE: With the U.S. export license to send GPUs to UAE, NVIDIA can directly supply infrastructure there. (Tom’s Hardware)

• Saudi Arabia / HUMAIN: Large‑scale GPU deployments to build AI factories in Saudi Arabia. (NVIDIA Newsroom)

• Japan / SoftBank: U.S. domestic scaling: NVIDIA has committed to building AI supercomputers and chip packaging/testing facilities in Arizona and Texas, increasing domestic production capacity. (CRN)

• China exposure under pressure: Though NVIDIA has historically earned substantial revenue from China, trade/export restrictions and customs crackdowns limit its ability to fully monetize there. (Financial Times)

☑️ #260 Sep 15, 2025

Mellanox Technologies Ltd.

reuters.com: [Excerpt] China says preliminary probe shows Nvidia violated anti-monopoly law.

ANTITRUST VIOLATIONS CAN CARRY HEFTY FINES

According to China’s antitrust law, companies can face fines of between 1% and 10% of their annual sales from the previous year. China generated $17 billion in revenue for Nvidia in the fiscal year ending January 26, or 13% of total sales, based on its latest annual report.

Nvidia's shares fell 2% in pre-market trading on Monday.

The Chinese market regulator's announcement comes as the U.S. and China hold trade talks in Madrid, where chips, including the ones made by Nvidia, are expected to be on the agenda.

The extent to which China can have access to cutting-edge AI chips is one of the biggest flashpoints in the U.S.-Sino war for tech supremacy.

+ Continue reading

+ Related content:

• scmp.com: [Excerpt] State Administration for Market Regulation

  国家市场监督管理总局

  The State Administration for Market Regulation (SAMR) is a prominent Beijing-based Chinese governmental body, established in 2018 to consolidate market supervision. Its primary mission is to enhance market oversight efficiency and maintain fair competition. SAMR’s main areas of activity include comprehensive market regulation, antitrust enforcement, product quality and safety, consumer protection, and intellectual property. As China’s primary antitrust regulator, it plays a pivotal role in standardising and maintaining market order, ensuring a trustworthy and competitive business environment across the People’s Republic of China.

☑️ #259 Sep 8, 2025

Build a real-time options monitoring dashboard for NVDA options

llmhub.dev: [Excerpt] Autonomous execution by LLMHub Agent.

Summary

Report: Real-time NVDA Options Monitoring Dashboard

Date: 2025-09-08

Prepared by: (automated task execution summary)

TL;DR

• Completed end-to-end prototype of a real-time NVDA options monitoring dashboard: selected a data provider (Alpha Vantage), implemented a FastAPI backend that polls and normalizes realtime options data and exposes REST + WebSocket endpoints, built a lightweight frontend that renders options chains with greeks/volume and highlights changes, and deployed/validated the system locally with continuous updates observed for >5 minutes.

• Current prototype uses Alpha Vantage’s REALTIMEOPTIONS demo API key (rate-limited, premium product) and a polling backend (POLLINTERVAL=60s). It is suitable for demonstration and local validation but requires a production API key, higher rate limits, and a few hardening / performance changes for production use.

1) Objective

Build a real-time options monitoring dashboard for NVDA (NVIDIA) options that:

• Ingests real-time NVDA options data,

• Normalizes and serves the data through low-latency REST and WebSocket endpoints,

• Presents an interactive frontend showing options chains, greeks, volume, and real-time change highlighting,

• Deploy and validate locally (or to cloud) and verify continuous updates.

2) Completed tasks and details

Task A — Data provider selection and credentials

• Action taken:

  • Selected Alpha Vantage as the real-time options data provider.

  • Located and reviewed the REALTIME_OPTIONS documentation: https://www.alphavantage.co/documentation/#realtime-options.

  • Tested with demo API key (”demo”) using sample NVDA queries.

• Findings:

  • The demo key works for examples and basic testing but is heavily rate-limited (typical free limits are in the order of ~5 requests/min and ~500/day — confirm current limits on Alpha Vantage site).

  • REALTIME_OPTIONS is marked a premium product on Alpha Vantage. Production access, higher rate limits, and full pricing require signup/purchase via their dashboard or sales channel.

• Action items:

  • Sign up for a paid/premium Alpha Vantage tier (or alternative provider) for production rate limits and commercial usage.

  • Store the production API key securely (environment variables, secrets manager).

Task B — Backend: FastAPI service (poll + normalize + REST + WebSocket)

• Implemented:

  • FastAPI backend that:

    • Polls Alpha Vantage REALTIME_OPTIONS for NVDA at configured intervals (ENV var POLL_INTERVAL=60s used in the run).

    • Normalizes returned option items to a consistent internal JSON format (strike, expiration, type, bid/ask/last, volume, open interest, greeks where provided).

    • Exposes REST endpoint: GET http://127.0.0.1:8200/options/latest which returns the latest normalized snapshot.

    • Exposes WebSocket endpoint: ws://127.0.0.1:8200/ws that pushes updates to connected clients whenever a new poll yields changes.

  • Repository / files created/updated:

    • main.py (and a backup main.py.bak)- requirements.txt- README.md

• Runtime details:

  • Service started in background, PID 13316.

  • Listening on 127.0.0.1:8200.

  • Server logs written to /tmp/nvda-service.log.

  • Environment: POLL_INTERVAL=60, demo API key used for validation.

• Issues encountered:

  • FastAPI DeprecationWarning about on_event usage during development iterations.

  • Transient “address already in use” bind errors while iterating on ports. Resolved by stopping earlier instances and settling on port 8200.

• Status:

  • REST and WebSocket endpoints respond and deliver live NVDA options updates during verification runs.

Task C — Frontend dashboard (static files + websocket client)

• Implemented:

  • Lightweight frontend with files:

    • index.html

    • style.css

    • script.js

  • Frontend behaviors:

    • Fetches /options/latest to get initial snapshot.

    • Connects to ws://127.0.0.1:8200/ws for incremental updates.

    • Renders calls and puts with greeks, volume, last/bid/ask.

    • Highlights cells/rows when values change in real-time.

  • Location: /home/desktop/Desktop/nvda-frontend/

  • Static server(s) started; recorded PIDs [13731, 13815, 13858] (multiple attempts/instances observed).

• Issues encountered:

  • curl http://127.0.0.1:9000/index.html returned “Empty reply from server” indicating the static server did not start cleanly on one attempt.

  • Multiple background http.server instances indicated stray servers were running; these were stopped before a clean static server was started.

• Resolution:

  • After stopping stray instances and starting the intended static server, the frontend was served and connected to the backend WebSocket as part of verification.

Task D — Deploy and validate locally, verify continuous updates

• Deployed endpoints (local):

  • Frontend: http://127.0.0.1:9000/index.html

  • REST: http://127.0.0.1:8200/options/latest

  • WebSocket: ws://127.0.0.1:8200/ws

• Validation:

  • Observed continuous NVDA options updates for over five minutes.

  • WebSocket clients logged per-minute update messages to /tmp/nvda-ws.log and /tmp/nvda-ws2.log.

  • Captured sample REST JSON payloads showing normalized option chain data.

  • Minor operational action: stopped stray static servers, then re-started the intended static server; no functional issues prevented validation.

• Status:

  • End-to-end prototype verified locally and continuously updating while polling Alpha Vantage with demo key (POLL_INTERVAL=60).

3) Key findings and limitations

• Data provider constraints:

  • Alpha Vantage REALTIME_OPTIONS is marked premium; the demo key works for testing but is rate-limited and not suitable for production traffic.

  • Poll-based ingestion from a rate-limited API restricts achievable update frequency and concurrency. If sub-minute latency or many simultaneous clients are required, a premium/push/streaming product or higher rate limit is necessary.

• Backend design considerations:

  • Current model polls at fixed interval (60s) and broadcasts changes via WebSocket; this is simple, reliable for a prototype, and low-overhead but introduces poll latency equal to the interval.

  • WebSocket pushes deliver low-latency updates to connected clients between polls.

• Operational notes:

  • FastAPI on_event DeprecationWarning observed — review FastAPI lifecycle hooks if upgrading to newer framework versions.

  • Careful port/process management is necessary when iterating locally; multiple http.server instances caused an “Empty reply” failure until cleaned up.

• Security and configuration:

  • Demo API key in use — do not embed demo key in production assets or source control.

  • No TLS on the local WS/HTTP endpoints in the prototype. Production deployment should use TLS (wss/https) and secure API key storage.

• Performance:

  • POLL_INTERVAL=60 is conservative; reducing it requires more API quota or streaming. For many clients, consider central broadcast (single poll + server-side fanout) rather than each client polling the provider.

• Observability:

  • Logs were written to /tmp/nvda-service.log and WebSocket logs to /tmp/nvda-ws*.log. Add metrics (request latency, poll durations, message rates) and structured logs for production.

4) Files, endpoints and runtime artifacts (observed)

• Backend:

  • Files: main.py (and main.py.bak), requirements.txt, README.md

  • Running process: PID 13316

  • Address: 127.0.0.1:8200

  • Logs: /tmp/nvda-service.log

  • REST endpoint: GET /options/latest

  • WebSocket endpoint: /ws

  • Environment used: POLL_INTERVAL=60, demo API key

• Frontend:

  • Files: index.html, style.css, script.js

  • Location: /home/desktop/Desktop/nvda-frontend/

  • Static server attempts: PIDs [13731, 13815, 13858]

  • Local UI URL used for validation: http://127.0.0.1:9000/index.html

  • Observed issue: initial static server returned “Empty reply from server” before cleanup

• Validation logs:

  • WebSocket logs: /tmp/nvda-ws.log, /tmp/nvda-ws2.log (per-minute update messages)

• REST payload samples captured during validation (stored where indicated in README or logs)

5) Recommendations and next steps (to reach production readiness)

Short-term (near-term):

• Acquire a production/premium API key from Alpha Vantage (or an alternative provider) to obtain required rate limits and commercial terms.

• Move API key to environment variable or secret manager; remove any hard-coded demo keys from code or public repos.

• Harden the backend:

  • Implement exponential backoff and jitter for provider errors / rate-limiting responses.

  • Implement request rate limiting and queuing to avoid provider throttles.

  • Add health endpoints, readiness/liveness checks, and structured logging.

  • Add metrics (Prometheus) for polls/sec, payload sizes, WS messages, latencies.

• Improve frontend serving:

  • Use a single managed static server or a simple reverse proxy (nginx) to avoid stray http.server instances.

  • Add CORS rules and appropriate security headers.

• Security:

  • Use TLS for all external-facing endpoints (https/wss).

  • Use token-based auth or other protections to limit access to endpoints.

• DevOps:

  • Containerize the app (Docker) and provide docker-compose for easy local dev.

  • Add CI checks and automatic linting.

  • Add an automated test harness to validate poll/update flow and edge cases.

Medium-term (performance and scale):

• Replace or augment polling with provider push/streaming if available to reduce latency and API quota usage.

• Add a caching/broadcast layer (Redis Pub/Sub or Kafka) for durability and scale; backend polls once and fanouts to many WebSocket connections.

• Implement client reconnection/back-off logic and delta compression for updates to reduce bandwidth when many clients connect.

• Evaluate alternative data providers that offer low-latency streaming for options, especially for high-frequency requirements.

Suggested verification and testing

• Load test WebSocket fanout with realistic client counts to verify server and network limits.

• Simulate provider rate-limiting/failures and verify backend backoff and client behavior.

• Latency testing: measure time from provider publish -> backend poll -> client receive for different poll intervals and under load.

6) How to reproduce locally (quick)

• Prereqs: Python 3.10+, pip install -r requirements.txt

• Set environment:

  • export ALPHAVANTAGEAPIKEY=”demo” (replace with production key when available)

  • export POLL_INTERVAL=60

• Start backend:

  • uvicorn main:app --host 127.0.0.1 --port 8200 (or use the provided run script)

• Start static frontend:

  • cd /home/desktop/Desktop/nvda-frontend

  • python -m http.server 9000 --bind 127.0.0.1

• Open: http://127.0.0.1:9000/index.html

• Observe:

  • REST: GET http://127.0.0.1:8200/options/latest

  • WS: ws://127.0.0.1:8200/ws

• Logs: /tmp/nvda-service.log, /tmp/nvda-ws.log, /tmp/nvda-ws2.log

7) Final comments

The project delivered a functional local prototype for real-time NVDA options monitoring with REST and WebSocket APIs and a responsive frontend that highlights live changes. The approach is fit for demonstration and local validation. The primary blocker for production readiness is provider access (premium API/streaming and quota limits), along with standard production hardening (secure keys, TLS, metrics, robust backoff). With a production data feed and the additions recommended above, this architecture can be scaled to support multiple symbols, higher update rates, and many simultaneous clients.

If you’d like, I can:

• Produce a prioritized implementation plan and estimated effort to move to production (including cost estimates for Alpha Vantage or alternative feeds).

• Create Dockerfiles and a docker-compose manifest for reproducible local deploys.

• Draft code changes to implement backoff, caching, and metrics in the FastAPI backend.

+ Related content:

• PrateekJ17 (update; 10/12/25): So we started this because we needed a way to provision infrastructure for AI agents at scale. Turns out nobody had really solved this properly. We wanted agents that could just... spin up a VM when they needed one, do their thing, and shut down. Simple idea, but getting it to actually work reliably was a whole journey. What we ended up building:

Agents that can deploy and control their own virtual machines An orchestration layer that doesn’t fall over when you scale to 100+ agents Support for pretty much any LLM (GPT-5, Claude, local models, whatever) Real monitoring because debugging invisible agents is a nightmare

• The whole thing is open source. Apache license. No strings attached. We’re using it for automated testing and web scraping stuff, but honestly people are probably going to use it for things we haven’t even thought of. If you’ve ever tried to run computer-use agents in production, you know the pain points. That’s what we tried to fix. GitHub: https://github.com/LLmHub-dev/open-computer-use

  • PrateekJ17 (update; 10/14/25): We open-sourced XAI’s Macrohard, an autonomous computer-using agent.

    • https://github.com/LLmHub-dev/open-computer-use

☑️ #258 Sep 5, 2025

GAIN AI Act

@nvidianewsroom

+ Related content:

• aimagazine.com (9/9/25): [Excerpt] What is the GAIN AI Act and Why Does Nvidia Oppose it?

  The GAIN AI Act is introduced as part of US export controls requiring Nvidia and other domestic AI chip makers to prioritise American buyers.

  The new rules would potentially apply even to older AI GPUs like Nvidia’s HGX H20 or L2 PCIe models, which were specifically designed to comply with previous export restrictions while serving international markets.

☑️ #257 Sep 3, 2025

GPU market share

@firstadopter: GPU market share, according to Jon Peddie Research

• Nvidia 94%

• AMD 6%

+ Related content:

• jonpeddie.com > Nvidia’s new B30A chip (8/22/25): [Excerpt] Targeted for Chinese market amid regulatory constraints.

  Production of the B30A is expected to use TSMCs CoWoS-S packaging method. This design will accommodate one compute chiplet and four HBM3E stacks, lowering costs relative to the CoWoS-L packaging used in Nvidia’s dual-chiplet B200 and B300 designs. The simplified configuration supports streamlined manufacturing while maintaining performance appropriate for export compliance.

☑️ #256 Sep 2, 2025

Categorically false

@nvidianewsroom: We've seen erroneous chatter in the media claiming that NVIDIA is supply constrained and "sold out" of H100/H200.

As we noted at earnings, our cloud partners can rent every H100/H200 they have online — but that doesn't mean we're unable to fulfill new orders.

We have more than enough H100/H200 to satisfy every order without delay.

The rumor that H20 reduced our supply of either H100/H200 or Blackwell is also categorically false — selling H20 has no impact on our ability to supply other NVIDIA products.

​​

💲 EARNINGS ANNOUNCEMENT: Aug 27, 2025

NVIDIA Announces Financial Results for Second Quarter Fiscal 2026

• Revenue of $46.7 billion, up 6% from Q1 and up 56% from a year ago

• Data Center revenue of $41.1 billion, up 5% from Q1 and up 56% from a year ago

• Blackwell Data Center revenue grew 17% sequentially

⇢ Some reactions:

• 007ofWallSt: NVIDIA's $60B buyback is a classic signal of confidence—or desperation. Are they trying to prop up a sagging stock, or is this a strategic move to fend off dilution and attract more institutional interest? Either way, watch the order flow closely.

• @kakashiii111: Two direct customers of Nvidia were responsible for 44.4% of the total Q2's revenue of the Data Center. Read that again: HALF of Nvidia's Data Center revenue in Q2 comes from just TWO customers.

  

☑️ #255 Aug 25, 2025

To Robot: Enjoy your new brain

NVIDIA Jetson Thor

This powerful new robotics computer is designed to power the next generation of general and #HumanoidRobots in manufacturing, logistics, construction, healthcare, and beyond /@NVIDIARobotics/​​ 

+ Related content:

• nvidianews.nvidia.com: [Excerpt] NVIDIA Blackwell-Powered Jetson Thor Now Available, Accelerating the Age of General Robotics.

  NVIDIA today announced the general availability of the NVIDIA Jetson AGX Thor™ developer kit and production modules, powerful new robotics computers designed to power millions of robots across industries including manufacturing, logistics, transportation, healthcare, agriculture and retail.

  Early adopters include industry leaders Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic and Meta, while 1X, John Deere, OpenAI and Physical Intelligence are evaluating Jetson Thor to advance their physical AI capabilities.

  

  • blogs.nvidia.com: [Excerpt] NVIDIA Jetson Thor Unlocks Real-Time Reasoning for General Robotics and Physical AI.

    Agility Robotics, a leader in humanoid robotics, has integrated NVIDIA Jetson into the fifth generation of its robot, Digit — and plans to adopt Jetson Thor as the onboard compute platform for the sixth generation of Digit. This transition will enhance Digit’s real-time perception and decision-making capabilities, supporting increasingly complex AI skills and behaviors. Digit is commercially deployed and performs logistics tasks such as stacking, loading and palletizing in warehouse and manufacturing environments.

☑️ #256 Aug 22, 2025

Nvidia's Huang talks to TSMC, weighs up China chip demand

@ReutersBiz: WATCH: Nvidia CEO Jensen Huang arrives in Taipei to visit chip foundry partner TSMC, as the world's most valuable company navigates rising friction between Washington and Beijing over access to its industry-leading AI chips https://www.reuters.com/business/finance/nvidia-ceo-taipei-visit-tsmc-says-talks-with-us-over-new-china-chip-2025-08-22/

☑️ #255 Aug 22, 2025 🔴 rumor

H20 halted

@CryptoMilox: $NVDA has requested Foxconn to suspend work on its H20 AI chip, the most advanced model it can sell in China, according to Reuters. Foxconn was responsible for backend processing. Nvidia stated, “We constantly manage our supply chain to address market conditions.” #AI @grok

⚡️

@grok: Thanks for the tag! Confirmed via Reuters and other sources: Nvidia paused H20 chip production with Foxconn due to market conditions, amid US export limits and China's push for domestic AI tech. This could impact NVDA's China revenue, but demand for other chips remains strong.

+ Related content:

• theinformation.com: [Excerpt] Nvidia Orders Halt to H20 Production After China Directive Against Purchases.

  Nvidia has told some of its component suppliers to suspend production work related to the H20, its chip tailor-made for the Chinese market, according to two people with direct knowledge of the communications.

☑️ #254 Aug 19, 2025

Speculated B30A

@ohlennart: The speculated B30A would be a really good chip. “50% off” is false reassurance.

• ½ B300 performance, ½ price = same value (just buy 2x)

• Well above (12x!) export control thresholds

• Outperforms all Chinese chips

• Delivers 12.6x the training perf of the H20 -Better than H100

⚡️

[2] This is probably Nvidia's response to Trump's statement to “take 30% to 50% off of it.” Don't be fooled. This works for some products, but not for chips in an exponential world. It's well above all thresholds, better than the H100, and if half-priced, it might be as good.

[3] If it's half the performance but also half the cost of the B300, just buy two B30As? You get equivalent aggregate performance. This undermines export controls. It's probably just literally half of the B300: one logic die instead of two, with 4 HBM stacks instead of 8.

[4] Think of it this way: if I want to buy an acre of land and buy it as two plots of half an acre each, it doesn't change how much land I now own. Same logic applies here: half performance at half price means you just buy twice as many.

[5] Sure there's networking and other overhead, but it would be a small percentage of total cost of ownership, which China will be more than eager to pay.

[6] B30A sits well above all export control thresholds (12x above!), requiring a license that might get granted. But export controls should create a widening gap over time: exponentials work in our favor. At only 50% below and half cost, you just buy two.

[7] On pure specs alone, it crushes any Chinese chip. Plus it can be produced and sold at quantity while China struggles with scale. Factor in software ecosystem advantages (CUDA, reliability) and the gap widens further. 7/

[8] Compared to the H20 that just got approved, the B30A delivers 12.6x better FLOP/s performance, meaning 12.6x better for training AI systems. Don't get tricked. Selling this is contra to export control strategy and would be a strategic error.

+ Related content:

• reuters.com (5/26/25): [Excerpt] Exclusive: Nvidia to launch cheaper Blackwell AI chip for China after US export curbs, sources say.

  ANOTHER CHIP

  Nvidia's market share in China has plummeted from 95% before 2022, when U.S. export curbs that impacted its products began, to 50% currently, Nvidia CEO Jensen Huang told reporters in Taipei last week.

  Huang also warned that if U.S. export curbs continue, more Chinese customers will buy Huawei's chips.

  According to two of the sources, Nvidia is also developing another Blackwell-architecture chip for China that is set to begin production as early as September. Reuters was not immediately able to learn the specifications of that variant.

☑️ #253 Aug 18, 2025

More than 2 million developers now using the NVIDIA robotics stack

blogs.nvidia.com: [Excerpt] Celebrating More Than 2 Million Developers Embracing NVIDIA Robotics.

Peer Robotics, Serve Robotics, Carbon Robotics, Lucid Bots, Diligent Robotics and Dexmate are just a few of the companies making a splash. Stay tuned here for their stories — and some product news you won’t want to miss.

This week, we’re celebrating the more than 2 million developers now using the NVIDIA robotics stack. These builders are reshaping industries across manufacturing, food delivery, agriculture, healthcare, facilities maintenance and much more.

Since the launch of the NVIDIA Jetson platform in 2014, a growing ecosystem of more than 1,000 hardware systems, software and sensor partners have joined the thriving developer community to help enable more than 7,000 customers to adopt edge AI across industries.

+ Continue reading

+ Related content:

• @NVIDIAAIDev (update; 8/20/25): Half a million downloads? No big deal.

  #NVIDIACosmos Reason — an open, customizable, 7B-parameter VLM — is helping shape the future of physical AI & robotics.

  See why it's one of the top 10 most downloaded NVIDIA models on @huggingface right now https://huggingface.co/nvidia/Cosmos-Reason1-7B

  

☑️ #252 Aug 15, 2025

We write to urge you to reverse your decision to allow AMD and Nvidia to sell advanced AI semiconductor chips ("covered chips") to the People's Republic of China ("PRC"), in exchange for a fee

@AnthonyAdragna: News: Six senior Democrats — @MarkWarner, @SenSchumer, @SenJackReed, @SenatorShaheen, @SenWarren and @ChrisCoons — demand Trump reverse course on Nvidia chips deal. "These sales to a leading adversary run counter to U.S. national security interests." https://warner.senate.gov/public/_cache/files/3/a/3a2a1db4-65d2-4377-822a-5f245d53b28f/62DE3F69670AE9734557935EEC1CB6A941532F5BA23C1E543B74B4EECAD3D420.250815.trump-senate-15-percent-fee-letter.signed.pdf

[Excerpt] We again urge your Administration to quickly reverse course and abandon this reckless plan to trade away U.S. technology leadership. In order to assist the public in its understanding of this proposed deal, we request the following information by August 22:

1. What entities or parties were involved in the "negotiation" that resulted in the 15 percent payment requirement as a condition for granting a license to sell covered chips to the PRC?

2. What laws, regulations, or other factors were identified, raised, evaluated, and/or otherwise determined to be relevant in those negotiations, including any legal opinions provided by Administration attorneys used to inform the legality of various approached proposed during those deliberations?

3. Who was responsible for determining that this arrangement complies with United States laws, regulations, and international obligations?

4. Describe the manner in which the 15 percent fee will be determined and collected

5. What individual(s) or entities are responsible for making the determination and collection identified in item 4?

6. List all purposes to which the 15 percent fee payments will be directed once collected.

7. What other companies are under consideration for this kind of deal utilizing export control laws?

☑️ #251 Aug 14, 2025 🟠 opinion

Feel free to disagree with me, time will tell who’s right

@DarioCpx: I don’t think people realise the magnitude of what’s happening here

• 2 GPUs smugglers arrested by the FBI on US soil, 100% they will sing like birds in the hope for a shorter jail time. This means they will start incriminating people higher and higher in the scheme till the top is reached. Who sits at the very top? $NVDA

• Of course $NVDA will try to come up with a black sheep among its employees pretending a “rogue salesman” could smuggle billions of USD worth of GPUs without raising any suspect

• $SMCI will try to come up with a similar excuse pretending another “rouge employee” managed to control billions worth of USD of server racks used as a carrier to smuggle its GPUs without raising any suspect in the company

• Big problem for both $SMCI and $NVDA: by the look of it the FBI has been investigating since a long time and it might be hard to cover up all trails that lead high up to the companies management

• At this point the ball will be in the US government court: if they pull the plug on all of this a good chunk of the market will crash led by the duo $SMCI / $NVDA on the way down. If they decide to let this go and just punish some scapegoats, then the companies will have to choose between stopping or continuing. If they continue and this all racket becomes a matter of national security then the US government might well change his mind on the matter.

• China already figured out relying on $NVDA GPUs will be a matter of national security for itself too. Why? Best case scenario US forbids any selling and smuggling of GPUs to China hindering its AI sector growth. Worst case scenario $NVDA hid a kill switch in its GPUs and this gives the power to the US government to turn off China data centres at will. Clearly the only rational move from China is to replace $NVDA in its infrastructure as soon as possible even if this will temporarily slow the development of AI in the country.

As you can easily understand, no matter the scenario $NVDA is going to lose a huge chunk of revenues that US Big Tech aren’t eager to make up for. US hyperscalers are already in advanced development of their own chips to run AI in a more scalable and energy efficient way compared to using GPUs (that let’s not forget is a 20y+ old technology meant to serve gaming and graphic industry).

$NVDA will then try to use its own GPUs to run its own cloud infrastructure becoming effectively a competitor of the other Big Techs. It will have a very hard time to succeed here and sell enough services to make up from loss in revenues of selling GPUs - objectively speaking.

No matter the scenario, $NVDA is doomed to crash spectacularly in the future in a similar fashion it already happened TWICE after the DotCom bust and the 2018 crypto mining bust.

All I described doesn’t even take into account $NVDA round-tripping scheme to inflate its revenues, I assumed that all the sales it reported so far were real (that fyi is different from “legal”)

Feel free to disagree with me, time will tell who’s right.

⚡️

@DarioCpx (8/13/25): If this is in the news it means the FBI likely already did it and if that’s the case $NVDA but most of all $SMCI are in huge troubles and better they start lawyering up big time @kakashiii111

⚡️

@DarioCpx (8/6/25): I bet the FBI will raid $SMCI offices and open a hell of a pandora box… “Authorities [..] seized phones belonging to Geng and Yang and found evidence of communications about shipping chips covered by export controls to China through Malaysia in violation of US restrictions.”

🟩 SP5 > Nvidia > Latest >

​

☑️ #189 Nov 8, 2025

A new $38 billion partnership between AWS and OpenAI highlights Amazon ongoing challenges. The deal involves AI cloud servers that only contain Nvidia GPUs, with no mention of Trainium processors

@hypertechinvest: AI startup Cohere found that Amazon Trainium 1 and Trainium 2 chips were underperforming Nvidia H100 GPUs, according to an internal confidential Amazon document from July obtained by Business Insider.

Cohere reported that access to Trainium 2 was extremely limited and affected by frequent service disruptions, the document noted.

Stability AI, a startup known for generating AI images, had similar concerns. It concluded that Trainium 2 underperformed the H100 on latency, making it less competitive in both speed and cost.

An Amazon spokesperson said the company is grateful for customer feedback that helps make its chips even better and more widely used. The Cohere case is not current, the spokesperson added, noting that Trainium and Inferentia have achieved great results with customers including Ricoh, Datadog, and Metagenomi.

“We expect to accommodate more customers starting with Trainium 3, previewing later this year,” the spokesperson said. “We’ll accomplish that as we always do by listening to our customers and remaining vocally self-critical as we continue innovating to give more customers access to chips with the best possible price and performance.”

Last week, Amazon CEO Andy Jassy said during the company’s earnings call that Trainium 2 chips are fully subscribed and represent a multibillion-dollar business growing 150% quarter-over-quarter. Spokespeople for Cohere and Stability AI didn’t respond to requests for comment.

Other AWS customers have also raised complaints about Amazon AI chips. According to the July document, a startup called Typhoon found Nvidia A100 GPUs to be up to three times more cost-efficient than AWS Inferentia 2 chips for certain workloads.

Similarly, a research group called AI Singapore determined that AWS G6 servers, equipped with Nvidia GPUs, offered better cost performance than Inferentia 2 across multiple use cases. Inferentia chips are used for inference, while Trainium chips focus on training models.

A new $38 billion partnership between AWS and OpenAI highlights Amazon ongoing challenges. The deal involves AI cloud servers that only contain Nvidia GPUs, with no mention of Trainium processors.

☑️ #188 Oct 14, 2025

Amazon’s major recent developments

chat.openai.com (chatgpt) and phind.com (Phind-70B) > [Sources not verified. Do your own research]:

Major Recent Developments

• Shreveport Next‑Generation Fulfillment Center: Amazon has opened a large, highly automated fulfillment center in Shreveport, Louisiana (SHV1). This facility is powered by AI/robotics (including “Proteus,” a fully autonomous mobile robot), is significantly larger and more robot‑intensive than prior ones, and is expected to employ around 2,500 people at full capacity. (Amazon)

• Investments in Sustainability & Packaging: Amazon continues to push on environmental goals. For example, it has removed plastic air pillows from delivery packaging in its global fulfillment centers and retrofitted its packing machines in many U.S. centers to produce made‑to‑fit paper bags. (Amazon)

• Enhancements to Delivery Systems: The Vision‑Assisted Package Retrieval (VAPR) system is being introduced into electric vans, reducing physical/mental effort for delivery drivers and saving time per route. (Amazon)

• Expanding Pharmacy & Healthcare‑adjacent Services: Amazon is expanding Amazon Pharmacy’s same‑day delivery into more cities, embedding pharmacies within Same‑Day Delivery sites to enable rapid medication delivery. (Amazon)

• Combating Counterfeits & Strengthening Trust: The company released reports showing it identified and removed/seized large numbers of counterfeit items globally, blocked suspicious listings proactively, and invested over $1 billion in brand protection measures. (Industry Intelligence Inc.)

Strategic Partnerships

• GMG in MENA: Amazon has entered a strategic partnership with GMG (a retail/well‑being conglomerate) to enhance its digital retail capabilities in the Middle East & North Africa. The partnership covers fast delivery, product selection in categories like sports, health & beauty, and improving customer experience via Amazon’s technology (AI, logistics, customer service). (gmg.com)

• bpost in Belgium: Amazon is expanding its collaboration with Belgium’s national postal service (bpost) for last‑mile delivery as part of its infrastructure and service enhancements in Belgium. (corporate.flandersinvestmentandtrade.com)

Product Innovations

• AI & Robotics in Fulfillment: The Shreveport facility includes significantly more robotics (roughly ten times more than older fulfillment centers), new systems like Sequoia to manage very large inventories, and Proteus robots. (TechCrunch)

• Application of AI in Shopping & Logistics: Features like AI Shopping Guides in Amazon’s app, VAPR for drivers, tools helping sellers (e.g. Project Amelia) to manage listings, improvements in operations using ML and vision systems, etc. (Amazon)

• Sustainability as Product/Packaging Innovation: Transitioning packaging to more recyclable or renewable materials; reducing packaging weight; eliminating certain plastics. (Amazon)

Global Expansion

• Belgium: Amazon is investing over €1 billion between 2025‑2027 in Belgium to improve logistics, deliver same‑day options, strengthen local infrastructure including robotics and mechatronics, support SMEs, and partner with bpost. (Euro Weekly News)

• Middle East & North Africa (MENA): Through the GMG partnership, Amazon is expanding its e‑commerce capabilities and delivery speed there, enabling customers in regions like UAE, Saudi Arabia, Egypt to get more product categories via Amazon with faster delivery. (gmg.com)

• Seller tools / Global Selling: Amazon is increasing efforts to enable sellers globally (and in particular, Chinese sellers) to use its infrastructure to reach more international markets, simplifying cross‑border logistics, and expanding its site footprint to more countries. (equalocean.com)

☑️ #187 Sep 22, 2025

“Fake sales” (Prime Day)?

The case appears to still be in its early phases (“proposed class action”). 

chat.openai.com (chatgpt) > [Sources not verified. Do your own research] > What the lawsuit alleges:

• A proposed class‑action lawsuit filed on September 22, 2025 in Washington state (case Armstrong et al v. Amazon.com, Inc.) claims Amazon used inflated “List Prices” to compute percentage‑off discounts during Prime Day, making the deals appear much deeper than they really were. (classaction.org)

• The complaint says that for many items, the “List Price” (the strikethrough price shown before the discount) was never actually Amazon’s price for 90 days or more before the sale. (Top Class Actions)

• Examples cited include:

  • A pair of headphones listed at a “List Price” of $179.95 and shown as 44% off — but plaintiffs claim the real comparable price on Amazon was always in the $130‑$160 range. (CBS News)

  • A kids’ tablet claimed to be 40% off (list price $119.99, sale price $72.18) even though the tablet had been sold for ~$50–$85 in the 90 days prior. (CBS News)

• The suit argues that because of the misleading “discounts” consumers may have paid more than they would have (or foregone shopping around) based on the belief of getting a great deal. (classaction.org)

  • The lawsuit is separate from other legal actions against Amazon (for example the one by the Federal Trade Commission (FTC) over “dark patterns” in its Prime subscription service) which Amazon has settled. (TIME)

☑️ #186 Sep 3, 2025

Project Rainier: 1.3 gigawatts (GW) of datacenter capacity for Anthropic

@StockSavvyShay: $AMZN AWS is building over 1.3GW of datacenter capacity for Anthropic -- hosting nearly 1M Trainium2 chips.

Anthropic is now at a $5B run-rate (5x YoY) -- with AWS as its main training partner and inference largely on $GOOGL Cloud TPUs.

+ Related content:

• phind.com (Phind-70B): [Sources not verified. Do your own research] [Excerpt]: This infrastructure buildout represents one of the largest single-customer datacenter deployments in history.

  Technical Infrastructure:

  • The datacenters will be primarily equipped with AWS's custom Trainium chips

  • While Trainium2 trails Nvidia's GB200 in raw specifications:

    • FP16 FLOPs: 667 TFLOP/s (vs Nvidia's 2500 TFLOP/s)

    • Memory bandwidth: 2900GB/s (vs Nvidia's 8000GB/s)

    • Scale-up network bandwidth: 186 TB/s (vs Nvidia's 576 TB/s)

• semianalysis.com: [Excerpts] Amazon's AI Resurgence: AWS & Anthropic's Multi-Gigawatt Trainium Expansion.

  Anthropic multi-gigawatt clusters, Trainium ramp, best TCO per memory bandwidth, system-level roadmap, Bedrock and internal models.

  

  • Trainium vs GPUs

    While Amazon’s AI Datacenters are impressive in scale and speed, the design of individual building is unremarkable. Hyper-optimized for air-cooling, this blueprint is identical to 5-year-old traditional AWS Cloud datacenters.

    What makes these facilities unique is their inside: they’ll host the world’s largest cluster of non-Nvidia AI chips, with just under a million Trainium2 in the largest campus. To understand everything about the Trainium2 system, read our December 2024 technical deep dive.

    Trainium2 lags Nvidia’s systems in many ways, but it was pivotal to the multi-gigawatt AWS/Anthropic deal.

    

  ​

• aboutamazon.com (6/24/25): Meet Project Rainier, Amazon’s one-of-a-kind machine ushering in the next generation of AI.

  Our new AI “compute cluster” will connect hundreds of thousands of Trainium2 chips across the U.S.

  

☑️ #185 Aug 14, 2025

Expanding same-day grocery deliver

@joecarlsonshow: Amazon is up +3.5% today after news that they're aggressively expanding into grocery delivery. Grocery was the Achilles heel of Amazon, giving Walmart+ room to grow as people want fast and efficient grocery delivery. Now Amazon is aggressively expanding same-day grocery delivery to thousands of cities to stave off the expansion of Walmart+. I view this is a positive incremental development for Amazon, and I have more to say about it in an upcoming episode

☑️ #184 Aug 14, 2025

Trainium 3

@dnystedt: Amazon’s chip division is working with Taiwan’s Alchip on Trainium3 and Trainium4 to bring the chips to mass production on TSMC 3nm and 2nm processes, respectively, media report, citing unnamed industry sources. Trainium3 will be in mass production in the 1st quarter of 2026. Trainium4 will follow later, with Astera Labs on the I/O chips and Synopsys IP for 225G SerDes. 1/2 $AMZN $TSM $ALAB $SNPS $MRVL #Alchip #semiconductors #semiconductor http://ctee.com.tw/news/20250814700072-430501

+ Related content:

• aws.amazon.com: AI Accelerator - AWS Trainium

  The first-generation AWS Trainium chip powers Amazon Elastic Compute Cloud (Amazon... Trn1 instance features up to 16 Trainium chips and delivers up to 3 ...

☑️ #183 Jul 19, 2025

Insider trades

@statquants: $AMZN Amazon insiders (mostly from Jeff) have unloaded over $2.4B in shares in under 3 weeks heading into Q3 — ranking in the 92nd percentile, another record pace.

+ Related content:

• barchart.com (7/15/25): Amazon.com Inc (AMZN) - Insider Trades

  

☑️ #182 Jul 4, 2025 🔴 rumor

Amazon built a massive AI supercluster for Anthropic called Project Rainier – here's what we know so far

theregister.com: [Excerpt] It's almost like AWS is building its own Stargate.

Amazon Web Services (AWS) is in the process of building out a massive supercomputing cluster containing "hundreds of thousands" of accelerators that promises to give its model building buddies at Anthropic a leg up in the AI arms race.

The system, dubbed Project Rainier, is set to come online later this year with compute spanning multiple sites across the US. Gadi Hutt, the director of product and customer engineering at Amazon's Annapurna Labs, tells El Reg that one site in Indiana will span thirty datacenters at 200,000 square feet apiece. This facility alone was recently reported to consume upwards of 2.2 gigawatts of power.

But unlike OpenAI's Stargate, xAI's Collusus, or AWS's own Project Ceiba, this system isn't using GPUs. Instead, Project Rainier will represent the largest deployment of Amazon's Annapurna AI silicon ever.

"This is the first time we are building such a large-scale training cluster that will allow a customer, in this case Anthropic, to train a single model across all of that infrastructure," Hutt said. "The scale is really unprecedented."

+ Continue reading

+ Related content:

• ☑️ #165 Dec 3, 2024 AWS CEO Talks New Chip Clusters, Nvidia and AI Ambitions

  @markets: [Excerpt] AWS CEO Matt Garman discusses Amazon's new AI chip arrays and a large language model called Nova. Speaking on "Bloomberg The Close," Garman also comments on the company's partnerships with Intel and Nvidia. Amazon's new chip arrays, called Project Rainier, will contain hundreds of thousands of Trainium2 semiconductors, quintupling Anthropic's current processing power.

☑️ #181 Jun 29, 2025

Amazon (MONSTER) Insider Trading Alert

@Barchart: Founder Jeff Bezos to dump $5.4 BILLION worth of $AMZN shares

+ Related content:

• sec.gov > Form 4 [Statement of changes in beneficial ownership of securities] (update; 7/1/25)

  

  The first transactions took place on June 27 and June 30: Jeff Bezos sold more than 3,324,926 shares of the company generating about $736.7 million before taxes.

  • sec.gov > Form 144 [Report of proposed sale of securities] (6/27/25) > [Excerpt] Securities To Be Sold:

    • Title of the Class: Common

    • Date you Acquired: 07/05/1994

    • Nature of Acquisition Transaction: Purchase - Original Issue

    • Name of Person from Whom Acquired: Issuer

    • Is this a Gift? /

    • Date Donor Acquired: /

    • Amount of Securities Acquired: 25000000

    • Date of Payment: 07/05/1994

    • Nature of Payment *: Founder Stock

☑️ #180 Jun 4, 2025 🔴 rumor

Amazon is developing software for humanoid robots

reuters.com: [Excerpt] The company is developing the artificial intelligence software that would power such robots, the report said, adding that Amazon plans to use hardware from other firms in its tests, for now.

Amazon did not immediately respond to a request for a comment outside its regular business hours.

In a series of announcements on Wednesday, Amazon demonstrated how stockroom robots, delivery people and its sprawling warehouses will all benefit from a hefty dose of AI, speeding packages to customer doorsteps.

+ Continue reading

+ Related content:

• theinformation.com (6/4/25): Amazon Prepares to Test Humanoid Robots for Delivering Package.

  Amazon is developing software for humanoid robots that could eventually take the jobs of delivery workers.

  The Information reported citing a person familiar with the matter. The Information is subscription-only publication. 

☑️ #179 Jun 2, 2025

The Times and Amazon Announce an A.I. Licensing Deal

In 2023, The Times sued OpenAI and Microsoft for copyright infringement. Now its editorial content will appear across Amazon platforms.

The New York Times Company has agreed to license its editorial content to Amazon for use in the tech giant’s artificial intelligence platforms, the company said on Thursday.

The multiyear agreement “will bring Times editorial content to a variety of Amazon customer experiences,” the news organization said in a statement. Besides news articles, the agreement encompasses material from NYT Cooking, The Times’s food and recipe site, and The Athletic, which focuses on sports.

This is The Times’s first licensing arrangement with a focus on generative A.I. technology.

+ Continue reading

☑️ #178 May30, 2025

Amazon Pledges to Continue Aggressive Data Center Expansion

bloomberg.com: [Excerpt] Amazon.com Inc.’s cloud services chief says the company is aggressively expanding server farms globally and looking to boost access to the latest artificial intelligence chips from Nvidia Corp.

Amazon Web Services, the world’s largest seller of rented computing power and data storage, opened a cluster of data centers in Mexico earlier this year, AWS Chief Executive Officer Matt Garman said Friday in an interview with Bloomberg Television. The cloud division is also building out new facilities in Chile, New Zealand, Saudi Arabia and Taiwan, Garman said.

+ Continue reading

+ Related content:

• aboutamazon.com (update; 6/9/25): [Excerpt] Amazon plans to invest $20 billion in Pennsylvania to expand cloud computing infrastructure and advance AI innovation.

  Amazon plans to invest at least $20 billion in Pennsylvania to expand its AI infrastructure, creating 1,250 new high-skilled jobs and supporting thousands more in the AWS data center supply chain.

  Since 2010, Amazon has invested $26 billion in Pennsylvania, contributing more than $26 billion to the Commonwealth’s GDP and supporting 27,000 full and part-time jobs, as well as 37,100 indirect jobs across various facilities.

• press.aboutamazon.com (update; 6/5/25): [Excerpt] Amazon Launches Infrastructure Region in Taiwan.

  The AWS Asia Pacific (Taipei) Region consists of three Availability Zones at launch, giving AWS 117 Availability Zones across 37 AWS Regions globally.

  With today’s announcement, AWS has plans for 13 more Availability Zones and four more AWS Regions in Chile, New Zealand, the Kingdom of Saudi Arabia, and the AWS European Sovereign Cloud.

  The AWS Asia Pacific (Taipei) Region is sovereign-by-design, just as the AWS Cloud has been since day one.

  • aws.amazon.com: [Excerpt] AWS Global Infrastructure. The most secure, extensive, and reliable Global Cloud Infrastructure, for all your applications.

    AWS Global Infrastructure Map

    The AWS Cloud spans 117 Availability Zones within 37 geographic regions, with announced plans for 13 more Availability Zones and four more AWS Regions in New Zealand, the Kingdom of Saudi Arabia, Chile, and the AWS European Sovereign Cloud.

• aboutamazon.com (update; 6/4/25): [Excerpt] Amazon plans to invest $10 billion in North Carolina to expand cloud computing infrastructure and advance AI innovation.

  Amazon plans to invest $10 billion in North Carolina to expand its AI infrastructure, creating 500 new high-skilled jobs and supporting thousands more in the AWS data center supply chain.

  Since 2010, Amazon has invested $12 billion in North Carolina, contributing $13.1 billion to the state's GDP and supporting 24,000 full and part-time jobs plus 26,000 indirect jobs across various facilities.

☑️ #177 May 6, 2025 🟥 0-shot prompting

[TOPIC] ZOOX’s purpose-built autonomous vehicle: Scale up self-driving robotaxi production

phind.com (Phind-70B) & perplexity.ai: [Sources not verified. Do your own research] Recent software recall affecting 270 vehicles due to prediction accuracy problems. April 2025 collision in Las Vegas highlighted system limitations and engineering costs currently dominating operational expenses.

While Zoox aims to scale up production for US expansion, several factors indicate a gradual rather than rapid transition to full commercialization. Zoox's path forward appears to be characterized by careful, incremental progress driven by technological advancement, regulatory compliance, and market conditions.

The company's position under Amazon's backing provides stability for this measured approach, allowing for necessary investment in solving complex technical challenges rather than rushing toward arbitrary deadlines.

+ Related content:

• aboutamazon.com: ZOOX

• zoox.com > @ZooxYouTube (4/3/25): Steering the Future | Episode 1.

• perplexity.ai: Zoox’s inaccurately confident prediction explained. What happened?

  Zoox, Amazon’s autonomous vehicle subsidiary, issued a voluntary recall of 270 robotaxis (2024 Zoox sedans) after an incident in Las Vegas on April 8, 2025. The root cause was a software defect: at speeds over 40 mph, the Zoox vehicle could make an “inaccurately confident prediction” about how another vehicle would behave, especially when a car entered its lane from a perpendicular driveway.

  • static.nhtsa.gov (5/1/25; pdf): [Excerpts] Part 573 Safety Recall Report 25E-029

    • Description of Defect: In situations where Zoox vehicles are operating at speeds greater than 40mph, and a vehicle incrementally encroaches from a perpendicular driveway approaching the Zoox vehicle's lane and subsequently halts, the Zoox vehicle may make an inaccurately confident prediction that the vehicle is going to proceed into the Zoox vehicle's lane. The Zoox vehicle may plan a path of travel based on the inaccurate prediction of the other vehicle trajectory, resulting in the inability to respond in a timely fashion to avoid a collision.

    • Chronology: On April 8, 2025, in Las Vegas, NV, an unoccupied Zoox robotaxi operating in autonomy was involved in a collision with a passenger car. There were no injuries. Zoox paused all driverless vehicle operations across all geofences to conduct an internal review.

      On April 10, 2025, Zoox confirmed, with simulation and analysis, that this condition would occur only at speeds over 40mph. Based on this analysis, driverless vehicle operations were unpaused in geofences with maximum speeds below 40 mph, but excluded Las Vegas routes above 40 mph.

      On April 11, 2025, after further analysis, Zoox confirmed the root cause, implemented a software update, and began testing it both in simulation and on drivered, retrofitted test vehicles.

      On April 13, 2025, Zoox reported the incident to NHTSA in accordance with the Standing General Order (SGO

      Report #30610-10473-1).

      On April 15, 2025 Zoox verified the software update through the safety clearance process.

      On April 16 and 17, 2025, updated software was released to all impacted Zoox vehicles operating on public roads. Following confirmation of the deployment of the updated software, operations resumed in all geofences.

      The Zoox Safety Committee met on April 17 and 23, 2025 and after final analysis and review of all the facts, made a determination to voluntarily and proactively submit this Part 573 notification for the previously remedied software.

☑️ #176 May 1-2, 2025

Jeff Bezos plans to sell $4.75 billion worth of Amazon stock

ir.aboutamazon.com > SEC > 8-K > 10-Q: Amazon said in its quarterly report filed with the SEC that Jeff Bezos adopted a trading plan on March 4 "intended to satisfy Rule 10b5-1(c)" to sell up to 25 million of the company's common stock over a period ending on May 29, 2026.

• This aligns with a plan announced in May 2024, set to complete by year-end, under the SEC’s 10b5-1 rule to avoid insider trading issues.

  • Form 144 (7/2/24): Report of proposed sale of securities

    • Number of Shares or Other Units To Be Sold: 25,000,000

    • Number of Shares or Other Units Outstanding: 10,406,627,415

☑️ #175 Apr 30, 2025

The United States of Amazon

Kyla’s Newsletter: maga-maoism, long degeneracy, GDP.

Kyla’s Newsletter

The United States of Amazon

Good morning from Stanford! We have just a few stickers left from the Sticker drop - if you already ordered they should be on the way to you now! The funds are going directly to help me buy some filming gear for a project I am working on around economics education…

Read more

a year ago · 310 likes · 49 comments · kyla scanlon

Kyla’s Newsletter

human-centric economic analysis to help all of us understand the world better

By kyla scanlon

☑️ #174 Apr 29, 2025

Amazon to display tariff costs for consumers

punchbowl.news: [Excerpt] Amazon doesn’t want to shoulder the blame for the cost of President Donald Trump’s trade war.

So the e-commerce giant will soon show how much Trump’s tariffs are adding to the price of each product, according to a person familiar with the plan.

The shopping site will display how much of an item’s cost is derived from tariffs – right next to the product’s total listed price.

+ Continue reading

+ Related content:

• cnbc.com: [Excerpt] Amazon says displaying tariff cost ‘not going to happen’ after White House blowback.

  “The team that runs our ultra low cost Amazon Haul store considered the idea of listing import charges on certain products,” Amazon spokesperson Tim Doyle said in a statement. “This was never approved and is not going to happen.”

• @WhiteHouse: The right way → @Axios: 'Walmart is expanding support for American-made products through new programs to help small businesses.'

  THE TRUMP EFFECT.

  

☑️ #173 Apr 10, 2025

CEO Andy Jassy’s 2024 Letter to Shareholders

aboutamazon.com: [Excerpt] Next generation Whys.

While the team and I feel quite optimistic about the progress and potential of our existing businesses, we have plenty of new Whys we’re asking. Below are a few of them and some quick thoughts.

Why is AI so important? Will it really have as much impact as some claim and when?

Generative AI is going to reinvent virtually every customer experience we know, and enable altogether new ones about which we’ve only fantasized. The early AI workloads being deployed focus on productivity and cost avoidance (e.g. customer service, business process orchestration, workflow, translation, etc.). This is saving companies a lot of money. Increasingly, you’ll see AI change the norms in coding, search, shopping, personal assistants, primary care, cancer and drug research, biology, robotics, space, financial services, neighborhood networks—everything. Some of these areas are already seeing rapid progress; others are still in their infancy. But, if your customer experiences aren’t planning to leverage these intelligent models, their ability to query giant corpuses of data and quickly find your needle in the haystack, their ability to keep getting smarter with more feedback and data, and their future agentic capabilities, you will not be competitive. How soon? It won’t all happen in a year or two, but, it won’t take ten either. It’s moving faster than almost anything technology has ever seen.

OK, I buy AI is big; but why invest this much this quickly?

Fundamentally, if your mission is to make customers’ lives better and easier every day, and you believe every customer experience will be reinvented by AI, you’re going to invest deeply and broadly in AI. That’s why there are more than 1,000 GenAI applications being built across Amazon, aiming to meaningfully change customer experiences in shopping, coding, personal assistants, streaming video and music, advertising, healthcare, reading, and home devices, to name a few. It’s also why AWS is quickly developing the key primitives (or building blocks) for AI development, such as custom silicon AI chips in Amazon Trainium to provide better price-performance on training and inference, highly flexible model-building and inference services in Amazon SageMaker and Amazon Bedrock, our own frontier models in Amazon Nova to provide lower cost and latency for customers’ applications, and agent creation and management capabilities.

+ Continue reading

☑️ #172 Apr 2, 2025

Here’s what to expect from Project Kuiper’s first full-scale satellite launch

aboutamazon.com: [Excerpt] United Launch Alliance will send 27 Kuiper satellites into low Earth orbit as we begin a full-scale deployment of our satellite internet network.

+ Continue reading

+ Related content:

• @ulalaunch (update; 4/10/25):The launch of a United Launch Alliance Atlas V rocket carrying the Kuiper 1 mission for the Amazon’s Project Kuiper was scrubbed today due to inclement weather. A new launch date will be announced when approved on the range.

☑️ #171 Feb 27, 2024

Amazon Web Services announces a new quantum computing chip

aboutamazon.com: [Excerpt] New 'Ocelot' chip uses scalable architecture for reducing error correction by up to 90% and accelerating the development of real-world quantum computing applications.

New approach to quantum error correction

To address the current problems associated with quantum error correction, researchers at AWS developed Ocelot. Ocelot was designed from the ground up with error correction “built in.” “We looked at how others were approaching quantum error correction and decided to take a different path,” said Painter. “We didn't take an existing architecture and then try to incorporate error correction afterwards. We selected our qubit and architecture with quantum error correction as the top requirement. We believe that if we're going to make practical quantum computers, quantum error correction needs to come first.” In fact, according to Painter, his team estimates that scaling Ocelot to a “fully-fledged quantum computer capable of transformative societal impact would require as little as one-tenth of the resources associated with standard quantum error correcting approaches.”

AWS’s reverse logistics program provides a sustainable path for components—all while maintaining the highest level of data security.

One way to think about quantum correction is in the context of quality control in manufacturing, and the difference between needing one inspection point to catch all defects, instead of 10 inspection points. In other words, it offers the same result, but with fewer resources and an overall improved manufacturing process. By reducing the amount of resources needed through approaches such as with Ocelot, quantum computers can be built smaller, more reliably, and at lower cost. All of this accelerates the path to applying quantum computing to future applications in the real-world, such as faster drug discovery and development, the production of new materials, the ability to make more accurate predictions about risk and investment strategies in financial markets, and many more.

+ Continue reading

+ Related content:

• @AmazonNews (2/25/25): Take a look inside AWS's quantum computing lab.

• California Institute of Technology (Caltech) > Campus Map > Amazon Web Services (AWS) Center for Quantum Computing

  

☑️ #170 Dec 20, 2024

It’s a Bird! It’s a Plane! It’s Brazilian Bum Bum Cream!

nytimes.com: [Excerpt] Recently, Amazon has developed a new drone — known as the MK-30 — that it claims solves many of the problems with previous models. The new drones fly twice as far, and the company says they are significantly quieter. The drones can also identify where to drop packages with the help of cameras and sensors, which means no more yard surveys or QR codes.

This year, Amazon began offering drone delivery from a facility in Tolleson, Ariz., a suburb of Phoenix — where, as in Texas, the skies are typically sunny and clear, making for optimal flying conditions. Its drones now deliver dozens of packages a day to real, paying customers in the West Valley Phoenix Metro area — a tiny fraction of what even one Amazon van could do, but a step toward fulfilling Mr. Bezos’ original vision.

+ Continue reading

☑️ #169 Dec 9, 2024

Intuit and Amazon Expand Strategic Partnership

investors.intuit.com: [Excerpt] Intuit QuickBooks will become Amazon’s preferred partner for financial management solutions integrated directly into Amazon Seller Central to fuel the success of Amazon sellers.

MOUNTAIN VIEW, Calif.--(BUSINESS WIRE)-- Intuit Inc. (NASDAQ: INTU) and Amazon (NASDAQ: AMZN) today announced a multi-year strategic partnership to empower millions of Amazon sellers to manage their finances, stay compliant, access capital, and grow their business. By leveraging Intuit’s AI-driven expert platform, millions of Amazon sellers will be able to discover and access Intuit’s platform seamlessly, benefiting from powerful financial insights like profitability, cash flow, and estimated tax liabilities to fuel their growth.

Key benefits for sellers include:

• Comprehensive financial management.

• Simplified access to capital.

+ Continue reading

☑️ #168 Dec 9, 2024

Bitcoin Treasury Assesment

@FreeEntProject: The rumors are true: #Microsoft won't be our last #Bitcoin proposal.

On Friday we presented a Bitcoin proposal to #Amazon. Read our $AMZN proposal here: https://nationalcenter.org/wp-content/uploads/2024/12/Amazon-Proposal.pdf

$BTC $MSFT #Crypto #CryptoNews #cryptocurrency

☑️ #167 Dec 6, 2024

Limits of automation in Shreveport (SHV1)

wsj.com: [Excerpt] Amazon’s New Robotic Warehouse Will Rely Heavily on Human Workers.

The 3 million-square-foot distribution center in Shreveport, La., demonstrates the limits of automation in big logistics operations.

Amazon just opened its most-automated warehouse yet. But underneath the robotics and artificial-intelligence technology at the site, the facility will still rely on thousands of employees.

The e-commerce behemoth said its 3 million-square-foot building in Shreveport, La., is its first warehouse to use automation and artificial intelligence at every step of the fulfillment process.

+ Continue reading

+ Related content:

• aboutamazon.com (10/9/24): Amazon unveils the next generation of fulfillment centers powered by AI and 10 times more robotics.

  

  ​

• nlep.org (10/26/23): Amazon Invests in Second Facility in Shreveport.

  This facility will be the second Amazon location in Shreveport, with the first being a Robotics Fulfillment Center slated to open in mid-2024. At Robotics Fulfillment Centers, employees work side-by-side with robots to pick, pack, and ship customer orders such as books, toys, and housewares. A last-mile facility is where customer orders are prepared for delivery to customers.​

☑️ #166 Dec 4, 2024 🟠 opinion

Limits of automation in Shreveport (SHV1)

@RihardJarc: Investors have underestimated $AMZN's AI strategy with the narrative that $AMZN is behind in AI.

Now, this narrative is proving out to be wrong, and the stock is getting rerated as:

• $AMZN now has a significant stake in Anthropic (the number 2 LLM provider right now), and AWS becoming Anthropic's main cloud provider.

• AWS is seeing big AI adoption with their Bedrock open strategy approach, with the AWS CEO saying yesterday that the number of users on AWS Bedrock in the last year alone had 5x.

• Jassy saying that you will see a 25% lower cost to serve this holiday because of AI and robotics

• $AMZN making significant strides with its custom silicon AI ASIC development with companies like Anthropic using AWS custom chips, but also $AAPL becoming their chip client.

Once again, the market's lack of understanding of a company's fundamental development and domain knowledge is the reason why there are always opportunities out there. As Philip Fisher said it well: The stock market is filled with individuals who know the price of everything, but the value of nothing.

☑️ #165 Dec 3, 2024

AWS CEO Talks New Chip Clusters, Nvidia and AI Ambitions

@markets: [Excerpt] AWS CEO Matt Garman discusses Amazon's new AI chip arrays and a large language model called Nova. Speaking on "Bloomberg The Close," Garman also comments on the company's partnerships with Intel and Nvidia. Amazon's new chip arrays, called Project Rainier, will contain hundreds of thousands of Trainium2 semiconductors, quintupling Anthropic's current processing power.

☑️ #164 Dec 3, 2024

PwC, AWS Expand Strategic Alliance to Catalyze Generative-AI powered Transformation for Industry Customers

pwc.com: [Excerpt] Collaboration will leverage new Amazon Bedrock Guardrails Automated Reasoning checks, announced today at AWS re:Invent.

According to the PwC US 2024 Cloud and AI Business Survey, companies that are seeing increased productivity and creating new revenue streams with GenAI are more than twice as likely to realize these benefits through industry-specific solutions. Together, PwC and AWS will build a series of industry-specific applications based on proven use cases leveraging the broad selection of proven foundation models from leading AI companies like Amazon—including Amazon Nova—a new generation of foundation models (FMs) that have state-of-the-art intelligence across a wide range of tasks, and industry-leading price performance. Amazon Bedrock also comprises the new Automated Reasoning checks, the first and only GenAI safeguard that helps prevent factual errors due to hallucinations, using logically accurate and verifiable reasoning. This gives customers confidence to more easily deploy powerful and explainable GenAI solutions that can enhance decision-making and accelerate digital transformation.

PwC will also integrate Amazon Bedrock Automated Reasoning checks within existing industry solutions, such as the Maestro Medical Legal Review solution, to validate that LLM-generated content complies with regulations, and within call center solutions, helping customers more easily confirm that agents comply with respective company protocols.

+ Continue reading

☑️ #163 Dec 3, 2024

Introducing Amazon Nova, our new generation of foundation models

aboutamazon.com: [Excerpt] New state-of-the-art foundation models from Amazon deliver frontier intelligence and industry-leading price performance.

As the next step in our AI journey, we’ve built Amazon Nova, a new generation of foundation models (FMs). With the ability to process text, image, and video as prompts, customers can use Amazon Nova-powered generative AI applications to understand videos, charts, and documents, or generate videos and other multimedia content.

+ Continue reading

+ Related content:

• AWS re:Invent updates

• https://x.com/_philschmid/status/1864016010464080260

☑️ #162 Dec 3, 2024

Chip Trainium2 AI para Ultraclusters

press.aboutamazon.com: [Excerpt] AWS Trainium2 Instances Now Generally Available.

LAS VEGAS--(BUSINESS WIRE)-- At AWS re:Invent, Amazon Web Services, Inc. (AWS), an Amazon.com, Inc. company (NASDAQ: AMZN), today announced the general availability of AWS Trainium2-powered Amazon Elastic Compute Cloud (Amazon EC2) instances, introduced new Trn2 UltraServers, enabling customers to train and deploy today’s latest AI models as well as future large language models (LLM) and foundation models (FM) with exceptional levels of performance and cost efficiency, and unveiled next-generation Trainium3 chips.

• New Amazon EC2 Trn2 instances, featuring AWS’s newest Trainium2 AI chip, offer 30-40% better price performance than the current generation of GPU-based EC2 instances.

• New Trn2 UltraServers use ultra-fast NeuronLink interconnect to connect four Trn2 servers together into one giant server, enabling the fastest training and inference on AWS for the world’s largest models.

+ Continue reading

+ Related content:

• AWS re:Invent updates

  

• aws.amazon.com: Why Trainium?

  AWS Trainium chips are a family of AI chips purpose built by AWS for AI training and inference to deliver high performance while reducing costs.

  The first-generation AWS Trainium chip powers Amazon Elastic Compute Cloud (Amazon EC2) Trn1 instances, which have up to 50% lower training costs than comparable Amazon EC2 instances. Many customers, including Databricks, Ricoh, NinjaTech AI, and Arcee AI, are realizing performance and cost benefits of Trn1 instances. Read more

Ultracluster in 2025: A massive Al supercomputer powered by its Trainium chips offering up to 83.2 petaflops of computing power. 

☑️ #161 Nov 27, 2024

Annapurna Labs

siliconangle.com: [Excerpt] Amazon’s secretive AI weapon: An exclusive look inside AWS’ Annapurna Labs chip operation.

Nestled on the ninth floor of an unassuming building in an affluent area of Austin, Texas dubbed The Domain, is one of the world’s most influential research labs powering modern artificial intelligence.

SiliconANGLE got an exclusive, behind-the-scenes tour of Amazon Web Services Inc.’s Annapurna Labs, where the cloud powerhouse does the secretive work of designing and testing the upcoming generations of itsAI accelerator Trainium and custom cloud compute Graviton chips.

Upon arrival, Martin was greeted in the lobby and escorted to the Labs, a highly restricted area even for Amazon.com Inc. employees. After being checked in, given a temporary visitor’s pass and a handful of Jolly Ranchers, Martin met with Rami Sinno, director of silicon engineering at AWS. Joining AWS from Arm Holdings PLC, Sinno brought large-scale project management experience to the project since its inception, detailing the challenging work of recruiting a team of talented engineers for the Trainium chip, a completely stealth initiative at the time.

+ Continue reading

☑️ #160 Nov 26, 2024

Japan competition authorities raid Amazon Japan, source says

asia.nikkei.com: [Excerpt] TOKYO (Reuters) -- Japan's Fair Trade Commission on Tuesday conducted an on-site inspection of Amazon Japan on suspicion of violating anti-monopoly laws, a government source said.

The unit of Amazon.com is suspected of inappropriately urging sellers to lower their prices on its e-commerce site in exchange for giving their products advantageous placement on the site, the source said.

+ Continue reading

Amazon Japan was previously raided on similar suspicions in March 2018.

☑️ #159 Nov 25, 2024 🟠 opinion

Researchers identified censorship blocklists used by Amazon US

@citizenlab: NEW REPORT: Our latest report reveals that the US storefront of @amazon uses a system to restrict shipments of certain products. We found 17k+ products that were restricted from being shipped to specific world regions, with the most common type of product being books. #censorship #AmazonCensorship #BannedBooks https://citizenlab.ca/2024/11/analysis-of-censorship-on-amazon-com/

⚡️

@citizenlab: The books that were banned were largely related to sensitive topics like erotica, religion, the occult, and LGBTIQ issues. Even though these books were in stock and a seller was willing to ship, Amazon still restricted users from shipping them to that region. Countries affected are the UAE, Saudi Arabia, other Middle Eastern countries as well as Brunei Darussalam, Papua New Guinea, Seychelles, and Zambia.

⚡️

@citizenlab: However, there are some miscategorized books that fall victim to this censorship. Researchers identified censorship blocklists used by Amazon US that resulted in overly broad or miscategorized restriction of unrelated products. Some examples include books on breast cancer, Nietzsche’s Gay Science, and “rainbow” Mentos candy. Also banned are recipe books that invoke “food porn” euphemisms.

⚡️

@citizenlab: The study analyzes the complex censorship system #Amazon uses to restrict shipping . Our analysis reveals Amazon uses a model that is sensitive to the presence of keywords, (e.g., terms like “gay,” “porn”) but this method is flawed and flags books for keywords even when the content is not related to controversial topics.

⚡️

@citizenlab: The report concludes that @amazon fails to meet its obligations under the United Nations Guiding Principles on Business and Human Rights which guarantees the right to access information and ideas without censorship.

☑️ #158 Nov 23, 2024

Temu vs. Amazon

Bristlemoon Capital: Temu vs. Amazon is classic counter-positioning. $AMZN can’t compete head on with Temu because it’s hamstrung by its fee structure - its merchants simply can’t make money on ultracheap goods given the hefty fees Amazon charges. We explore this in our recent $PDD report.

☑️ #157 Nov 22, 2024

Amazon's total investment in Anthropic to $8 billion

anthropic.com: [Excerpts] Powering the next generation of AI development with AWS.

Today we’re announcing an expansion of our collaboration with Amazon Web Services (AWS), deepening our work together to develop and deploy advanced AI systems. This expanded partnership includes a new $4 billion investment from Amazon and establishes AWS as our primary cloud and training partner. This will bring Amazon's total investment in Anthropic to $8 billion, while maintaining their position as a minority investor.

+ Continue reading

☑️ #156 Nov 21, 2024

The AI Effect: Amazon Sees Nearly 1 Billion Cyber Threats a Day

wsj.com: [Excerpt] The technology has spawned a surge in hacking attempts, says cyber chief CJ Moses, while Amazon is also using it to powerfully amp up its threat-analysis capability.

Just as criminals have embraced AI, Amazon has also turned to the technology to drastically scale up its threat-intelligence capabilities. The company, given its vast presence online, can now view activity on around 25% of all IP addresses on the internet, it says, between its Amazon Web Services platform, its Project Kuiper satellite program and its other businesses, giving the company a sweeping view of hacker capabilities and techniques.

Amazon’s chief information security officer, CJ Moses, spoke with WSJ Pro Cybersecurity on how the company is approaching threat intelligence in the AI era. Prior to his current role, Moses ran security for Amazon Web Services, its cloud business, and before that investigated cybercrime at both the Federal Bureau of Investigation and the Air Force Office of Special Investigations.

+ Continue reading

☑️ #155 Nov 19, 2024

AI's Amazon moment

Mike’s Substack: Reinventing the $60T service economy.

Mike’s Substack

AI's Amazon moment

Not long ago, you could build a successful professional service business as a solo practitioner. A local accountant could serve dozens of small businesses. An independent insurance agent could support hundreds of families…

Read more

2 years ago · Mike Wadhera

Mike’s Substack

Mike Wadhera's Substack

By Mike Wadhera

☑️ #154 Nov 11, 2024

Nam3L3ss

techcrunch.com: [Excerpt] Amazon confirms employee data stolen after hacker claims MOVEit breach.

The confirmation comes after a threat actor claimed to have published data stolen from Amazon on notorious hacking site BreachForums. The individual claims to have more than 2.8 million lines of data, which they say was stolen during last year’s mass-exploitation of MOVEit Transfer.

The threat actor, operating under the alias “Nam3L3ss” claims to have published data allegedly stolen from 25 major organizations, cybersecurity firm Hudson Rock reports.  

In a statement given to TechCrunch on Monday, Amazon spokesperson Adam Montgomery confirmed that employee information had been involved in a data breach.

+ Continue reading

+ Related content:

• community.progress.com (6/25/24): MOVEit Transfer Critical Security Alert Bulletin – June 2024 – (CVE-2024-5806)

  ​

• progress.com (7/5/23): MOVEit Transfer and MOVEit Cloud Vulnerability.

  ​

• socradar.io (update; 11/12/24): [Excerpt] In the manifesto we mentioned, which was released alongside the data leak, Nam3L3ss claims they do not actively hack organizations but “simply monitor the dark web and exposed online cloud services.”

  

  ​

• @fs0c131y (update; 11/12/24): Based on ChatGPT's analysis of messages from Nam3L3ss, the hacker who exploited the critical vulnerability in MOVEit could be either a native English speaker or a non-native speaker writing quickly and informally, without careful editing

  

☑️ #153 Nov 11, 2024

US % Market Share of Parcel Volume

@libertyrpf: US Delivery market. Look at Amazon Logistics go!

☑️ #152 Nov 6, 2024

47th President

@JeffBezos: Big congratulations to our 45th and now 47th President on an extraordinary political comeback and decisive victory. No nation has bigger opportunities. Wishing @realDonaldTrump all success in leading and uniting the America we all love.

☑️ #151 Nov 5, 2024

Amazon Maintains Dominant Lead in the Cloud Market

@StockMKTNewz: Jeff Bezos just filed for his sale of 1.07M shares of Amazon $AMZN stock at $200.07 per share bringing in ~$213.8 Million (Pre-Tax)

+ Related content:

• @unusual_whales: You can see all Amazon $AMZN insider transactions, including those of Jeff Bezos, on Unusual Whales. We recently updated our entire fundamental stock pages. Come check them out: https://unusualwhales.com/stock/AMZN/insiders

  • @unusual_whales (11/5/24): BREAKING: Jeff Bezos of Amazon, $AMZN, has sold over 1,000,000 $AMZN shares at $200.07 (for roughly $213 million).

  • @unusual_whales (11/2/24): Jeff Bezos of Amazon $AMZN sold 16,354,620 shares of $AMZN worth $3,048,501,168 today (yes, $3 billion!!!!!!!)

    

🟩 SP5 > Amazon > Latest >

​🟩 Silicon & Data: 1 > 2

​

☑️ #112 Nov 8, 2025

2M token context window

@brianroemmele: Boom!

Grok 4 Fast now has a huge 2M token context window!

Why is this important?

This context window can allow substantially documents or code bases to be loaded as a single prompt for interrogation and consideration.

With the added bonus of being insanely fast with almost no memory leakage.

I am running test this morning that are quite mind blowing. More soon.

⚡️

@brianroemmele: New Grok upgrade:

-grok-4-fast-reasoning: 77.5% -> 94.1%

-grok-4-fast-non-reasoning: 77.9 -> 97.9%

My tests show significant ability increases in reasoning for potential agentic work.

The fastest improving AI model in history… (Via @xlr8harder)

☑️ #111 Nov 8, 2025

Semiconductors: The most important sector in the world

@ozeco: One of the most important article I’ve ever written, a collab with my friend @Strategist, is the most comprehensive deep dive you will ever find into semiconductors, the most important sector in the world, and it’s completely free!

Listen to the podcast or read the article to become an expert on semiconductors and understand the big trends from AI to semi production, fabless companies to equipment makers.

Have you ever wanted to understand:

• How is the semi supply chain structured, why is it so globalized and what are the bottlenecks

• What are the innovations at the heart of this space

• How are chips made and who are the companies at the heart of this $600bn sector.

• The fabless and foundry business models

• The rise of China as a semiconductor giant

• How the US is reshoring the semiconductor production with hundreds of billions of $

• Investment implications

Crack The Market

Semiconductors: The most important sector in the world

Read more

a year ago · 42 likes · Ozeco and Strategist

Crack The Market

Institutional level research helping you own equities you understand and learn about trends and companies that will underpin the growth and value creation of the next decades.

By Ozeco

☑️ #110 Nov 8, 2025

Top Data Center stocks

@hypertechinvest: A few months ago, I changed my mind on IREN

Today, I still believe IREN is one of the best opportunities

2 GW of approved capacity in their Sweetwater data center is massive

A deal for that capacity could be an industry game changer

Even if they only pursued colocation, at 2 million of revenue per MW, it would mean 4 billion in new sales with a 90% potential EBITDA margin

At 20x EBITDA, Sweetwater alone would imply a 36 billion valuation potential, twice the current market cap

If you want to get the deepest insights into all the data center opportunities, check out my data center industry analysis and don’t invest blindly

Daniel Romero

Top Data Center Stocks

In this article, I’ll provide an industry overview of the AI data center companies and Bitcoin miners that are transitioning to HPC data centers, while also covering some of them individually. This article is part of a series…

Read more

10 months ago · 52 likes · 4 comments · Daniel Romero

Daniel Romero

A weekly newsletter by an equity analyst focused on finding high-growth stocks

☑️ #109 Nov 8, 2025

AI Economy map

@claritybriefing: I tried to make the most understandable of current technological, material and political influence, map of the AI revolution.

Here’s how I see it:

The map shows three industrial cores and the forces that connect them. They influence each other through constraints: political, technological, or material. Each arrow reflects a specific dependency that shapes the AI economy.

United States → Europe

• Political influence

The United States exerts political pressure on Europe through export regulations. Washington controls access to advanced GPUs, EDA software, and certain semiconductor technologies. Because ASML operates under Dutch jurisdiction but builds machines that use American components and intellectual property, the company falls under US export rules. When the US restricts China’s access to EUV or high-end DUV tools, Europe must align. This creates a one-directional dependency: Europe designs and builds the tools, but the United States decides who can receive them.

Mechanism: extraterritorial export law and licensing authority.

Effect: the US sets the boundaries of who Europe can sell advanced lithography to.

United States → Taiwan

• Political influence

The US influences Taiwan’s semiconductor ecosystem through security guarantees, export controls, and strategic incentives. TSMC’s decision to build fabs in Arizona is a direct result of American political pressure combined with industrial subsidies. The US relies on TSMC for the most advanced logic chips, and it uses political tools to keep part of this capacity on American soil to reduce risk.

Mechanism: CHIPS Act incentives, security policy, export restrictions. Effect: TSMC expands in the US under strategic pressure, not pure economics.

Taiwan → United States

• Technological influence

Taiwan holds the world’s most advanced manufacturing capability. Every major US tech company depends on TSMC’s nodes for performance: Nvidia, AMD, Apple, Google. The pace at which TSMC moves from one node to the next sets the ceiling for American AI growth. If TSMC slows, US compute slows.

Mechanism: advanced node availability (3 nm, 2 nm).

Effect: US AI progress is gated by Taiwanese manufacturing.

Europe → Taiwan

• Technological influence

Europe influences Taiwan through lithography. Without ASML’s EUV machines and Zeiss optics, TSMC cannot produce its leading-edge chips. This is not a symbolic dependency. It is physical and absolute. Every wafer that leaves TSMC’s 3 nm or 2 nm line carries ASML’s imprint.

Mechanism: EUV machinery, optical subsystems, precision tooling.

Effect: Europe decides the pace at which Taiwan can advance nodes.

5. China → Europe

• Material influence

China controls much of the global supply of critical materials used by European and Japanese semiconductor toolmakers: rare earths, gallium, germanium, high-purity chemicals. ASML, Zeiss, and the entire European optics chain depend on this upstream flow. Europe cannot cut China out without losing access to essential inputs for lithography production.

Mechanism: rare earth refining, chemical dominance.

Effect: Europe’s manufacturing capacity depends on Chinese materials.

6. China → Taiwan

• Material and volume influence

China is both a major supplier of raw materials and a major consumer of chips. TSMC’s older nodes (28 nm, 16 nm, 7 nm) serve large Chinese customers. China also exports gases, wafers, and refined metals that feed into Taiwanese production. Even under sanctions, material interdependence remains high.

Mechanism: materials, industrial gases, customer volume.

Effect: Taiwan’s fabrication ecosystem cannot fully detach from China.

7. United States → China

• Political constraint

The direction of the arrow is not a flow but a block. The US limits China’s access to advanced GPUs, chip design software, and key lithography tools. This slows China’s ability to reach cutting-edge nodes and forces it to accelerate domestic alternatives.

Mechanism: export bans, licensing limits, entity lists.

Effect: China’s technological progress is constrained by US policy.

8. China → United States

• Material influence

Despite political tension, the US remains dependent on China for a large portion of critical minerals and components. Even a fraction of disruption in Chinese supply chains can slow American manufacturing, data center construction, and chip packaging.

Mechanism: rare earths, graphite, minerals used in batteries and cooling systems.

Effect: Chinese materials underpin American hardware infrastructure.

The structural picture

• The United States controls political levers.

• Europe controls the lithography frontier.

• Taiwan controls advanced manufacturing.

• China controls the materials that feed the system.

No region is self-sufficient. Each holds one essential piece and lacks two others.

This is why the map looks like a triangle of constraints rather than an alliance network.

☑️ #108 Nov 3, 2025

Data Centers in Space​

Starcloud.com: Data centers in space

“The H-100 is about 100 times more powerful than any GPU computer that has been on orbit before,”

“It will be the first time that a terrestrial-grade data-center GPU will be flown and operated in orbit.”

Philip Johnston, CEO and cofounder of Starcloud, told IEEE Spectrum.

+ Continue reading

☑️ #107 Nov 3, 2025

“THE SINGULARITY IS NEAR: WHEN HUMANS TRANSCEND BIOLOGY”

@TOzgokmen: Which is these still companies exist today?

Why should NVDA have a 10-year lead?

Who knew about NVDA 10 years ago?

☑️ #106 Oct 29, 2025

American (and swedish) lithography

@OmerCheeema: I live in Eindhoven, the ASML town. I was chatting with an employee at a gym last night. He works on the mechanical engineering side of the wafer table for EUV machines. The job is to align wafers to the source and ready it for exposure. The alignment needs to be accurate at a micrometer level. And about 900 people at ASML work on it. Substrate, the startup, has 12 employees as seen on LinkedIn. ASML should offer them jobs.

<>

@BloombergTV: @BloombergTV: A secretive American startup says it has chipmaking technology to rival ASML and take on TSMC.

Bloomberg’s @EdLudlow breaks the story on Substrate’s particle accelerator-powered lithography machine and plans for an American foundry http://bloom.bg/48Q0LG7

• news.bloomberglaw.com (10/28/25): [Excerpt] Thiel-Backed Chip Startup Is Seeking to Crack ASML Dominance (1) A secretive American startup has emerged with aspirations to challenge two titans of the semiconductor industry: ASML Holding NV and Taiwan Semiconductor Manufacturing Co.

  The company, Substrate, has developed a new chipmaking machine that uses particle acceleration to handle lithography — the crucial process of etching microscopic circuitry onto silicon wafers. Claiming to have solved one of technology’s toughest problems, Substrate aims to compete with ASML, whose equipment is currently the only option for making advanced processors.

+ Related content:

• substrate.com & read our purpose: [Excerpt] Return to Solicon Valley.

  Substrate was founded to return the United States to dominance in semiconductor fabrication on three core beliefs:

  • The cost of semiconductor fabrication is out of control, and it will only continue to trend upward without radical new technological innovation

  • The United States is dangerously reliant on sources of geopolitical and supply chain risk that we do not directly control

  • The only chance the United States has to return to dominance in semiconductor production would be to create a new type of more vertically-integrated foundry, one that continues pushing Moore’s Law in both performance and cost

  To achieve this, Substrate is building next-generation semiconductor fabs to return America to dominance in semiconductor production and will use our technology—a new form of advanced X-ray lithography—to power them.

  

• linkedin.com: Atomic Layer Etching for Nano Device Fabrication

  We provide an ALE-based method of manufacturing nanostructures with characteristic size below 20nm.

  This is a new method of nanostructure fabrication using the atomic layer etching process, which is inherently a damage-free etch process. Owing to the etching process’s selectivity to inclined surfaces, which has been recently discovered, the inclined surfaces can be used as a mask and in this way walls of tapered structures, which can be readily fabricated by e.g., dry etching or epitaxial growth, will provide masking during the atomic layer etching process. This is why this process provides access to fabrication of extremely small structures in a very precise and efficient way.

  • alixlabs.com (10/7/25) [Excerpt] AlixLabs Secures Notice of Allowance for US Patent for Innovative Semiconductor Manufacturing Technology.

    Lund, Sweden – October 7, 2025 – AlixLabs, a pioneering Swedish semiconductor equipment startup, today announced that the United States Patent and Trademark Office (USPTO) has issued a Notice of Allowance for the continuation of its internally named “Tetris” patent. The new patent strengthens and broadens AlixLabs’ intellectual property portfolio, covering key innovations in selective side wall masking technology that integrate self-aligned double patterning (SADP) with the company’s proprietary Atomic Layer Etching-based Pitch Splitting (APS™) process.

    

☑️ #105 Oct 18, 2025

Data (digital) tasks: new data wave

@bearlyai: Uber will give its drivers in the US an option to make money by doing “digital tasks”. These short minute-long tasks can be done anytime including while idling for passengers:

• data-labelling (for AI training)

• uploading restaurant menus

• recording audio samples of themselves

• narrating scenarios in different languages

The opportunity is quite larges based on fact that two AI data-labelling startups (Scale AI, Surge AI) each reached ~$30B valuations.

Earlier in October, Uber — which has 1m+ drivers in the US — had acquired Belgian data-labelling startup Segments AI for an undisclosed amount.

☑️ #104 Oct 17, 2025

Scaling Compute as a Moat: Going from 0 to 200,000 GPUs in just over 200 days was practically unheard of. That puts xAI’s Colossus ahead of all but a few hyperscalers in terms of raw compute, making this infrastructure the central competitive moat for LLM training and inference at scale.

The core idea: Whoever owns (or controls) the most compute wins.

In only 122 days, xAI built its first huge data center, Colossus in Memphis, Tennessee, equipping it with 100,000 Nvidia GPUs. 92 days later, capacity was boosted to 200,000 units. Now, Antonio Gracias (Valor Equity Partners) anchors xAI Corporation’s plan to raise $20 billion for a deal using a special purpose vehicle (SPV) that would acquire semiconductors from Nvidia Corporation and then lease them to  Colossus 2. 

perplexity.ai and phind.com (Phind-70B) > [Sources not verified. Do your own research]: In the AI arms race, models matter—but scale, speed of iteration, and compute density are increasingly the differentiators. xAI clearly sees this and is making infrastructure its front line.

xAI’s rapid execution in scaling its compute infrastructure is one of the most striking industrial feats in recent tech history. Building and doubling capacity at Colossus in just over seven months illustrates the extreme speed and capital availability shaping the current AI arms race. A few key implications stand out:

• Vertical coordination and financing: By using a special purpose vehicle (SPV) backed by Antonio Gracias and Valor Equity Partners, xAI is effectively separating asset ownership (semiconductors) from operational use (Colossus 2). This structure minimizes balance sheet strain while ensuring access to scarce Nvidia GPUs—an ingenious financial engineering move under today’s hardware constraints.

  • xAI’s infrastructure strategy for its “Colossus 2” data center in Memphis, Tennessee, represents a significant departure from traditional AI development. The initial “Colossus 1” site already boasts over 200,000 Nvidia H100 GPUs. For “Colossus 2,” the focus is shifting to even more advanced hardware, with plans for 550,000 Nvidia GB200 and GB300 GPUs, aiming for an eventual total of 1 million GPUs within the entire Colossus ecosystem. Elon Musk has publicly stated an audacious goal for xAI to deploy 50 million “H100 equivalent” AI GPUs within the next five years. This scale is unprecedented, requiring a “gigawatt-scale” facility – one of the largest, if not the largest, AI-focused data centers globally, with xAI constructing its own dedicated power plant, Stateline Power, in Mississippi, to supply over 1 gigawatt by 2027. (wral news)

    

• Strategic positioning vs. hyperscalers: This model could allow xAI to scale faster than Big Tech peers locked into more traditional data center expansion timelines. If successful, it may pioneer a new hybrid approach—AI companies as “GPU tenants” rather than infrastructure owners.

• Market dynamics and geopolitical influence: The Memphis facility marks a strategic domestic placement, likely aligned with U.S. policy goals to repatriate critical computing capacity. It also signals Nvidia’s continuing dominance in AI training chips, though at the cost of creating systemic dependency between the two firms.

• Capital and risk outlook: Raising $20 billion through a single-purpose leasing vehicle signals investor appetite is still robust for foundation-model infrastructure, but it raises questions about long-term ROI and operational leverage—especially as new architectures (custom ASICs, emerging competitors) evolve.

Overall, this move looks like a high-risk, high-return bet designed to secure xAI’s independence in compute supply while cementing its position as a world-scale AI developer. It shows Elon Musk’s consistent playbook: integrate critical capabilities early and finance with partners willing to think in decades, not quarters.

+ Related content:

• @markets: Nvidia to Invest in Musk’s xAI.

  Elon Musk’s artificial intelligence startup xAI is raising more financing than initially planned, tapping backers including Nvidia to lift its ongoing funding round to $20 billion, according to people familiar with the matter. Nvidia is investing as much as $2 billion in the equity portion of the transaction, the people said. Bloomberg’s Annabelle Droulers reports.

• PrateekJ17 (10/12/25): So we started this because we needed a way to provision infrastructure for AI agents at scale. Turns out nobody had really solved this properly. We wanted agents that could just... spin up a VM when they needed one, do their thing, and shut down. Simple idea, but getting it to actually work reliably was a whole journey. What we ended up building:

  Agents that can deploy and control their own virtual machines An orchestration layer that doesn’t fall over when you scale to 100+ agents Support for pretty much any LLM (GPT-5, Claude, local models, whatever) Real monitoring because debugging invisible agents is a nightmare

  The whole thing is open source. Apache license. No strings attached. We’re using it for automated testing and web scraping stuff, but honestly people are probably going to use it for things we haven’t even thought of. If you’ve ever tried to run computer-use agents in production, you know the pain points. That’s what we tried to fix. GitHub: https://github.com/LLmHub-dev/open-computer-use

  • PrateekJ17 (10/14/25): We open-sourced XAI’s Macrohard, an autonomous computer-using agent.

    • https://github.com/LLmHub-dev/open-computer-use

☑️ #102 Oct 15, 2025

Sustainable and adaptive Scake Data Centers and Builld-to-Scale solutions for global hyperscale, AI/HPC, and enterprise customers

aligneddc.com: [Excerpt] AIP was founded by BlackRock, Global Infrastructure Partners (GIP), a part of BlackRock, MGX, Microsoft, and NVIDIA to expand capacity of AI infrastructure and help shape the future of AI-driven economic growth. Its financial anchor investors include the Kuwait Investment Authority and Temasek.

In less than a decade, Aligned has evolved into one of the largest and fastest growing data center companies globally. The Company designs, builds, and operates cutting-edge data campuses and data centers for the world’s premier hyperscalers, neocloud, and enterprise innovators. Aligned’s portfolio includes 50 campuses and more than 5 gigawatts of operational and planned capacity, including assets under development, primarily located in key Tier I digital gateway regions across the U.S. and Latin America including Northern Virginia, Chicago, Dallas, Ohio, Phoenix, Salt Lake City, Sao Paulo (Brazil), Queretaro (Mexico), and Santiago (Chile).

+ Continue reading

☑️ #101 Oct 14, 2025

Beginning in calendar Q3 2026, Oracle will be the first hyperscaler to offer a publicly available AI supercluster powered by 50,000 AMD Instinct MI450 Series GPUs

oracle.com: [Excerpt] Oracle and AMD Expand Partnership to Help Customers Achieve Next-Generation AI Scale.

AMD Instinct MI450 Series GPUs Coming to OCI

AMD Instinct MI450 Series GPU-powered shapes are designed to deliver high-performance, flexible cloud deployment options and provide extensive open-source support. This provides the ideal foundation for customers running today’s most advanced language models, generative AI, and high-performance computing workloads. With AMD Instinct MI450 Series GPUs on OCI, customers will be able to benefit from:

• Breakthrough compute and memory.

• AMD optimized “Helios” rack design.

• Powerful head node.

• DPU-accelerated converged networking.

• Scale-out networking for AI.

• Innovative UALink and UALoE fabric.

• Open-source AMD ROCm™ software stack.

• Advanced partitioning and virtualization.

To give customers that build, train, and inference AI at scale more choice, OCI also announced the general availability of OCI Compute with AMD Instinct MI355X GPUs. These will be available in the zettascale OCI Supercluster that can scale to 131,072 GPUs. AMD Instinct MI355X-powered shapes are designed with superior value, cloud flexibility, and open-source compatibility. Learn more here and here.

+ Continue reading

+ Related content:

• chat.openai.com: [Sources not verified. Do your own research] Key Specifications & Architecture

  • The MI450 Series are AMD’s next‑generation data center / AI training & inference GPUs, built on CDNA “Next” architecture (sometimes referred to CDNA 5 in reporting). They will be manufactured using TSMC’s 2 nm (N2) process technology, which gives performance/power/density advantages over previous nodes. (Datacenter Dynamics)

  • AMD is positioning variants within the MI400 family: for example, MI450X, aimed at large‑scale AI workloads (low‑precision / inference), and MI430X, for HPC/high‑precision workloads (FP32/FP64). These variants will have differentiated logic (dropping certain precision units in each to optimize die space).

    The MI450X is expected to lean heavily on Ethernet‑based “Ultra Ethernet” (scale‑out networking), rather than relying immediately on UALink / dedicated accelerator fabric switches, which may lag in availability. (Tom’s Hardware)

  • “Helios” Rack Design

    • AMD is developing a rack‑scale architecture called Helios, which couples MI450 GPUs with EPYC CPUs (codename “Venice”) and networking via AMD Pensando (codename “Vulcano”) to deliver large scale AI training + inference clusters.

    • A Helios rack holds 72 MI450 GPUs, is liquid cooled, and is designed for high performance density, power efficiency, and scalability.

    • The Helios rack design signals AMD’s push to provide end‑to‑end solutions (GPU + CPU + networking) and compete more directly with Nvidia, especially in hyperscale cloud / AI supercluster infrastructure. (AMD)

​🟩 Silicon & Data: 1 > 2

🔊 Silicon’s Playlist

Les chants magnétiques / Magnetic Fields

​

☑️ #11 Nov 8, 2025

Most hedge funds today are tactical users of risk rather than long term directional players

@ksidiii: This kind of thing does not carry the same weight it once did. Most hedge funds today are tactical users of risk rather than long term directional players. If you look at the data, you will see that flows rotate much faster than they did twenty years ago, with positions turning over multiple times in different directions within just a few days. Beyond that, the magnitude of impact has changed as well. In the 1990s, hedge funds were seen as the powerhouses of the market. Today, outside of the major multistrat firms, that flow no longer moves the broader market in the same way. Your local financial advisor at JP Morgan probably moves more flow than your favorite mainstream movie fund manager.

⚡️

@firstsquawk:HEDGE FUNDS SHORTING AT FASTEST PACE SINCE APRIL

HEDGE FUNDS HAVE RAMPED UP SHORT POSITIONS TO THE HIGHEST LEVEL SINCE APRIL, SIGNALING GROWING BEARISH SENTIMENT ACROSS EQUITY MARKETS AS INVESTORS HEDGE AGAINST ECONOMIC WEAKNESS, RISING RATES, AND EARNINGS PRESSURE AMID INCREASING MARKET VOLATILITY.

☑️ #10 Aug 25, 2025 🟡 emailed

Why Stocks Look Rougher Than Indexes

Hello Friend,

In Boston, summer weather can feel wildly unpredictable from one day to the next. A July afternoon may reach 90 °F (32 °C) with humid sunshine, followed the very next day by thunderstorms and temperatures 15 degrees cooler at 75 °F (24 °C). The average July high is about 82 °F (28 °C), but the swings around that number can make summer days feel anything but average.

Looking more broadly across averages in Boston, the trend is clear. Average temperatures in Massachusetts have increased by about 2.75 °F (1.5 °C) over the past 55 years, while Suffolk County, where Boston is located, has seen an even sharper rise of about 3.3 °F (1.8 °C).

However, if we step back to view global averages, the picture looks steadier, though with a consistent upward drift. Worldwide land and ocean surface temperatures have climbed roughly 2.0 °F (1.09 °C) from pre-industrial times to 2011–2020, and in 2024 global average temperatures reached about 2.6 °F (1.46 °C) above that baseline, the hottest year recorded so far.

This framing sets the stage for the paper Why is the volatility of single stocks so much rougher than that of the S&P500? The authors explain that just as local weather is spikier than global climate, the volatility of individual stocks is far rougher than that of the overall market index. By using a Nested log Stationary fractional Brownian Motion factor model, they show how aggregation smooths out idiosyncratic shocks and leaves a common market factor as the main driver.

In quantitative finance, understanding volatility structure is central to risk management, option pricing, and portfolio construction. If single stocks behave with “super-rough” volatility but indices are smoother, models must distinguish between idiosyncratic noise and common factors to avoid mispricing risk and underestimating systemic exposure.

More detailed insights from the paper are provided below. If you’re interested in learning more about how quantitative models help us make sense of complex market behavior, sign up for our fall courses, covering machine learning, stochastic calculus, and the core tools of modern quantitative finance.

+ Quantopian

☑️ #9 Aug 25, 2025

Jocko

@waronweakness: 99% of risks aren't real

+ Related content:

• jocko.com: [Excerpt] JOCKO WILLINK. RETIRED U.S. NAVY SEAL, PODCASTER, AUTHOR, LEADERSHIP INSTRUCTOR, ENTREPRENEUR.

  Jocko Willink is a retired U.S. Navy SEAL officer, co-author of the #1 New York Times bestseller Extreme Ownership: How U.S. Navy SEALs Lead and Win, Dichotomy of Leadership, host of the top-rated Jocko Podcast, and co-founder of Echelon Front, where he serves as Chief Executive Officer, leadership instructor, speaker, and strategic advisor.

  Jocko spent 20 years in the SEAL Teams, starting as an enlisted SEAL and rising through the ranks to become a SEAL officer. As commander of SEAL Team Three’s Task Unit Bruiser during the battle of Ramadi, he orchestrated SEAL operations that helped the “Ready First” Brigade of the U.S. Army’s First Armored Division bring stability to the violent, war-torn city. Task Unit Bruiser became the most highly decorated Special Operations Unit of the Iraq War.

  Jocko Willink returned from Iraq to serve as Officer-in-Charge of training for all West Coast SEAL Teams. There, he spearheaded the development of leadership training and personally instructed and mentored the next generation of SEAL leaders who have continued to perform with great success on the battlefield. Jocko Willink is the recipient of the Silver Star, the Bronze Star, and numerous other personal and unit awards.

  Upon retiring from the Navy, Jocko Willink co-founded Echelon Front, a premier leadership consulting company, where he teaches the leadership principles he learned on the battlefield to help others lead and win. Jocko also authored the Discipline Equals Freedom Field Manual, a New York Times Bestseller, Leadership Strategies and Tactics, Final Spin, and the best-selling Way of the Warrior Kid children’s book series.

☑️ #8 Jul 13, 2025

This is a structural reordering of how capital defines risk

@onechancefreedm: Since 2009, and accelerating post 2020, capital has increasingly flowed into monopolistic, balance sheet-dominant mega caps while structurally abandoning small caps. The Russell 2000 is filled with firms highly sensitive to credit conditions, labor costs, and policy risk, many are unprofitable or carry heavy debt loads. So when real rates stay elevated, labor is tight, and inflation remains sticky, they bleed out, while tech monopolies with global reach and AI leverage attract all the oxygen.

This divergence is also psychological. Mega caps are seen as “safe growth,” increasingly behaving like quasi sovereigns or synthetic duration assets, tools for passive exposure, not just businesses. Small caps, on the other hand, reflect the real economy. Their collapse relative to tech signals that the perceived strength of the market is resting on an ever. narrower foundation.

This is a structural reordering of how capital defines risk. And if the Fed eventually pivots hard? Small caps may rip on beta sensitivity. But absent that, this ratio chart is a window into a deeply bifurcated economy and market.

⚡️

BREAKING: Small Cap Stocks $IWM is underperforming Mega Cap Tech Stocks by the largest margin in history.

☑️ #7 Jul 7, 2025

The man who beat Vegas, cracked Wall Street, and lived life on his terms

@quantitativo: Ed Thorp didn’t just study risk… he mastered it. He built the first wearable computer (with Claude Shannon) to beat roulette. He created the first quant hedge fund, with 227 winning months out of 230. He was the first to apply the Kelly Criterion in real-world investing. He bought Berkshire at $900. He was Citadel’s first LP. And then… he walked away.

Why? Because he realized something most investors never do:

Time, not money, is the real compounder.

This podcast episode distills key lessons from A Man for All Markets… the host calls Thorp “the only humble trader on Earth.” :) in it, we find:

• How Thorp thought about edge, survival, and position sizing

• His simple systems for health, wealth, and life

• Why he chose peace over scale, and what that says about success

• The quote that guided his decisions: “If you do this, what do you want to happen? And what do you think will happen?”

Genius in simplicity. Rich without stress. A trader who looked younger at 84 than at 73. Thorp’s life is the blueprint for anyone chasing more than just money.

+ Related content:

• quantitativo.com: [Excerpt] Who am I?

  My name is Carlos. I'm an aerospace engineer who started a career at a famous Swiss bank, became an entrepreneur, and sold two companies. I also led the expansion of a tech unicorn into some countries around the globe.

  Several years ago, in my Master's degree, I had the opportunity to learn quantitative trading and machine learning applied to trading from the Managing Director of JPMorgan AI research. I then decided to apply what I learned to my own capital, which is what I've been doing ever since.

  Quant Trading Rules

  Quant trading strategies, insights, and ideas. Backtested, forward-tested, and live-traded.

  By Quantitativo

☑️ #6 Jul 7, 2025 | archives

Capriva.io [FAQs]: What is risk?

capriva.io: Is my risk limited?

Risk is impossible to limit/eliminate because no one knows what the future holds. Even if it’s impossible to rule anything out , it’s true that some things are more likely than others.

Think of a weatherman predicting 11 degrees temperature on a London spring day. Both 11 and 55 degrees are possible on a given day, but 11 or thereabouts is far likelier than 55. It’s a question of probabilities.

We manage your risk in a way such that your likely returns (positive or negative) match your expectations. To this end, we must i) explain risk as a measure of future uncertainty, ii) understand your tolerance of risk/uncertainty and iii) manage your investment within this tolerance.

Think of risk as a probability - the likelihood of a given return happening, over a certain time-horizon. We summarise this probability as Value at Risk (VaR) - an industry standard measure for risk.

To explain VaR - think of the Apple Stock. Should it behave going forward like it has in the past, then it’s Value at Risk is 10%, because:

1. 1 month from today,

2. It is 95% likely that,

3. The Apple stock will trade between + and - 10%.

This is the case because in the past, it stayed in that range 19 out of 20 months (e.g. 95% of months). This also implies: Apple gained 10% or more / lost 10% or more one month in 20 (5%). By the same measure, the FTSE 100 has 7% VaR - Apple is more volatile/riskier - than the FTSE 100.

We’ll assess your risk appetite in terms of % VaR - then manage your positions below maximum VaR - meaning the odds of your losing more than the VaR we share will be 5% or less. Note that we can’t limit how much you lose in those 5% of months.

Having said that, coming back to the weatherman, is that with a 6% VaR, a 15% loss is far less likely than a 7% loss. (One would expect a 15% loss with a 6% VaR about 1 month every 4500 years).

In short: one cannot eliminate risk, but we manage the likelihood of undesired losses to a threshold tolerable by you, which we’ll share with you before you invest.

⚡️

capriva.io: How does Capriva manage my risk?

Capriva manages risk 24/5 by measuring, and if necessary, reducing risk for:

Any position that risks more, or for longer, than compatible with your target VaR.

Changes in market circumstances - e.g. increased volatility etc. - could push your risk above the maximum.

☑️ #5 Jun 25, 2025

Great models fail with bad data. At AIG, we had models that didn’t account for liquidity—and it nearly sank the system

databento.com: [Excerpt] Quants worth following: William Mann, Founder and CEO of HarmoniQ Insights.

What was the first sign that AIG’s super senior CDS exposure might be at risk?

“I can actually tell you the exact moment. I was sitting next to my boss on the trading floor. He turned to me and said, 'There’s a collateral call on the super senior.' We just stared at each other.”

At AIG Financial Products (AIG FP), Bill worked on a team that wrote credit default swaps (CDS) on “super senior” tranches of collateralized debt obligations (CDOs)—positions so far up the capital stack that over 95% of the underlying assets would have to default before any losses were realized.

“There was never, in any model, any expectation that you’d suffer a loss. They were designed for regulatory arbitrage.”

These trades were considered safe, essentially risk-free from a credit perspective. But that assumption didn’t account for mark-to-market risk.

“We’d been arguing with accountants for years about whether we should mark these to market. The models said they were safe. But when counterparties started marking them down and calling for cash, everything changed.”

Though no credit events had occurred, AIG’s contracts required collateral to be posted against unrealized losses. As markets seized up, counterparties exercised those rights.

By 2008, AIG had posted $33 billion in collateral. The CDS exposures were still rated investment grade, but the liquidity wasn’t. The risk wasn’t credit. It was confidence.

+ Continue reading

+ Related content:

• @Databento (6/26/25): Quants worth following: William Mann, Founder and CEO of HarmoninQ Insights.

• harmoniqinsights.com

☑️ #4 Mar 31, 2025

[Series: VaR] Testimony of Richard Bookstaber

bookstaber.com (9/10/09; pdf): [Excerpt] Submitted to the U. S. House of Representatives, Committee on Science and Technology Subcommittee on Investigations and Oversight.

For the Hearing:

"The Risks of Financial Modeling: VaR and the Economic Meltdown"

September 10, 2009

Mr. Chairman and members of the Committee, I thank you for the opportunity to testify today. My name is Richard Bookstaber. Over the past decade I have worked as the risk manager in two of the world's largest hedge funds, Moore Capital Management and, most recently, Bridgewater Associates. In the 1990s I oversaw firm-wide risk at Salomon Brothers, which at the time was the largest risk-taking firm in the world, and before that was in charge of market risk at Morgan Stanley.

I am the author of A Demon of Our Own Design - Markets, Hedge Funds, and the Perils of Financial Innovation. Published in April, 2007, this book warned of the potential for financial crisis resulting from the growth of leverage and the proliferation of derivatives and other innovative products.

Although I have extensive experience on both the buy-side and sell-side, I left my position at Bridgewater Associates at the end of 2008, and come before the Committee in an unaffiliated capacity, representing no industry interests.

My testimony will discuss what VaR is, how it can be used and more importantly, how it can be misused. I will focus on the limitations of VaR in measuring crisis risk. I will then discuss the role of VaR in the recent market meltdown, concluding with suggestions for ways to fill the gaps left by the limitations of VaR.

What is VaR?

VaR, or value-at-risk, measures the risk of a portfolio of assets by estimating the probability that a given loss might occur. For example, the dollar VaR for a particular portfolio might be expressed as "there is a ten percent probability that this portfolio will lose more than $VaR over the next day".

Here is a simplified version of the steps in constructing a VaR estimate for the potential loss at the ten percent level:

1. Identify all of the positions held by the portfolio.

2. Get the daily returns for each of these positions for the past 250 trading days (about a one-year period).

3. Use those returns to construct the return to the overall portfolio for each day over the last 250 trading days.

4. Order the returns for those days from the highest to the lowest, and pick the return for the day that is the 25*" worst day's return. That will be a raw estimate of the daily VaR at the ten percent level.
   Smooth the results by fitting this set of returns to the Normal distribution function.1

Limitations of VaR

The critical assumptions behind the construction of VaR are made clear by the process described above:

1. All of the portfolio positions are included.

2. The sample history is a reasonable representation of what things will look like going forward.

3. The Normal distribution function is a reasonable representation of the statistical distribution underlying the returns.

The limitations to VaR boil down to issues with these three assumptions, assumptions that are often violated, leading VaR estimates to be misleading.

Incomplete positions

Obviously, risk cannot be fully represented if not all of the risky positions are included in the analysis. But for larger institutions, it is commonplace for this to occur. Positions might be excluded because they are held off-balance sheet, beyond the purview of those doing the risk analysis; they might be in complex instruments that have not been sufficiently modeled or that are difficult to include in the position database; or they might be in new products that have not yet been included in the risk process. In the recent crisis, some banks failed to include positions in collateralized debt obligations (CDOs) for all three of these reasons?2 And that exclusion was not considered an immediate concern because they were believed to be low risk, having attained a AAA rating.

The inability to include all of the positions in the VaR risk analysis, the most rudimentary step for VaR to be useful, is pervasive among the larger institutions in the industry. This provides a compelling reason to have a 'flight to simplicity' in financial products, to move away from complex and customized innovative products and toward standardization.3

+ Continue reading (pdf)

+ Related content:

• moorecap.com: [Excerpt] Moore Capital Management, LP (“Moore Capital Management”) is a Delaware limited partnership that, with its predecessor firm, has been in the investment management business since 1989. Moore Capital Management is principally owned indirectly by Louis Moore Bacon. Mr. Bacon also serves as Chairman and Chief Executive Officer of Moore Capital Management. Moore Capital Management, together with its affiliates maintains offices in New York, London, Hong Kong and Florida.

• bridgewater.com: [Excerpt] What we do.

  Bridgewater Associates is a premier asset management firm, focused on delivering unique insight and partnership for the most sophisticated global institutional investors.

  Our investment process is driven by a tireless pursuit to understand how the world’s markets and economies work — using cutting edge technology to validate and execute on timeless and universal investment principles.

  Founded in 1975, we are a community of independent thinkers who share a commitment for excellence. By fostering a culture of openness, transparency, and inclusion, we strive to unlock the most complex questions in investment strategy, management, and corporate culture.

☑️ #3 Oct 15, 2024

Risk management

@theoldeconomy: A reasonably good analyst with excellent risk management skills will beat a brilliant analyst who is a mediocre risk manager.

Risk management aims first to ensure our survival in the markets and only then to give us the chance to make profits.

Even flawless risk management does not guarantee profits; it only increases the chances of achieving them.

☑️ #2 Oct 11, 2024

Why Your Risk Calculations Are Lying to You

DualEdge’s Substack: Why Relying on Beta and Volatility Is Setting You Up for Failure.

DualEdge’s Substack

Why Your Risk Calculations Are Lying to You

The real trouble with this world of ours is not that it is an unreasonable world, nor even that it is a reasonable one…

Read more

2 years ago · 7 likes · DualEdge Invest

DualEdge’s Substack

My personal Substack

By DualEdge Invest

☑️ #1 Sep 12, 2024

How to Think About Risk with Howard Marks

@oaktreecapital: Oaktree co-chairman Howard Marks explores the true meaning of risk in a new ten-part video course. He discusses the nature of risk, the relationship between risk and return, misconceptions about risk, and much more.

🔹Related content:

• Oaktreecapital.com > Archive Memos:

  • 2006: Risk

  • 2014: Risk revisited

  • 2015: Risk revisited again

• Key Takeaways

  • Risk is not volatility; “Risk is the probability of loss.”

  • Superior investors have a better sense of probability distributions and balance potential gains with risk control.

  • Continuous risk management is crucial, not just during perceived “risk-on” or “risk-off” markets.

  • Risk is counterintuitive and perverse. For example, assets become less risky when prices fall.

  • Riskier assets don’t necessarily produce higher returns.

Born in 1946, Howard Marks co-founded Oaktree Capital Management in 1995. Oaktree is now the largest investor in distressed securities worldwide. 

☑️ #0 Aug 5, 2024 | archives

The Price of Risk: With Equity Risk Premiums, Caveat Emptor!

Musing on Markets:

*Disclaimer4

🟩 SP5 > Google > Latest >

​

☑️ #212 Nov 6, 2025

With Ironwood, we can scale up to 9,216 chips in a superpod linked with breakthrough Inter-Chip Interconnect (ICI) networking at 9.6 Tb/s

@rohanpaul_ai: WOW. Google is finally rolling out its most powerful Ironwood AI chip, first introduced in April, taking aim at Nvidia in the coming weeks.

Its 4x faster than its predecessor, allowing more than 9K TPUs connected in a single pod.

The big deal is that many are not realizing, if it can do an all-reduce across 9k TPUs, it can run MUCH larger models than the Nvidia NVL72.

It would make really big 10T param size models like GPT-4.5 feasible to run. It’d make 100T param size models possible.

This is the last big push to prove that scaling works. If Google trains a 100T size model and demonstrates more intelligence and more emergent behavior, the AGI race kicks into another gear. If 100T scale models just plateau, then the AI bubble pops.

---

Ironwood TPUs go GA and Axion Arm VMs expand, promising upto 10X gains and cheaper inference while adding flexible CPU options for the boring but vital parts of AI apps.

Ironwood targets both training and serving with 10X peak over v5p and 4X better per chip than v6e for training and inference, so the same hardware can train a frontier model and then serve it at scale.

A single superpod stitches 9,216 chips with 9.6 Tb/s Inter-Chip Interconnect and exposes 1.77PB of shared HBM, which cuts cross device data stalls that usually throttle large models.

Optical Circuit Switching reroutes around failures in real time and the Jupiter fabric links pods into clusters with hundreds of thousands of TPUs, so uptime and scale do not fight each other.

On software, Cluster Director in GKE improves topology aware scheduling, MaxText adds straightforward paths for SFT and GRPO, vLLM runs on TPUs with small config tweaks, and GKE Inference Gateway cuts TTFT by up to 96% and serving cost by up to 30%.

Axion fills the general compute side of the house so data prep, ingestion, microservices, and databases stay fast and cheap while accelerators focus on model math.

N4A, now in preview, offers up to 64 vCPUs, 512GB memory, 50Gbps networking, custom machine types, and Hyperdisk Balanced and Throughput, aiming for about 2x better price performance than similar x86 VMs.

C4A Metal, in preview soon, gives dedicated Arm bare metal with up to 96 vCPUs, 768GB memory, Hyperdisk, and up to 100Gbps networking for hypervisors and native Arm development.

C4A provides steady high performance with up to 72 vCPUs, 576GB memory, 100Gbps Tier 1 networking, Titanium SSD up to 6TB local capacity, advanced maintenance controls, and support for Hyperdisk Balanced, Throughput, and Extreme.

⚡️

sundarpichai: Our 7th gen TPU Ironwood is coming to GA!

It’s our most powerful TPU yet: 10X peak performance improvement vs. TPU v5p, and more than 4X better performance per chip for both training + inference workloads vs. TPU v6e (Trillium). We use TPUs to train + serve our own frontier models, including Gemini, and we’re excited to make the latest generation available to @googlecloud customers.

+ Related content:

• cloud.google.com: [Excerpt] Announcing Ironwood TPUs General Availability and new Axion VMs to power the age of inference.

  TPUs are a key component of AI Hypercomputer, our integrated supercomputing system that brings together compute, networking, storage, and software to improve system-level performance and efficiency. At the macro level, according to a recent IDC report, AI Hypercomputer customers achieved on average 353% three-year ROI, 28% lower IT costs, and 55% more efficient IT teams.

  Ironwood TPUs will help customers push the limits of scale and efficiency even further. When you deploy TPUs, the system connects each individual chip to each other, creating a pod — allowing the interconnected TPUs to work as a single unit. With Ironwood, we can scale up to 9,216 chips in a superpod linked with breakthrough Inter-Chip Interconnect (ICI) networking at 9.6 Tb/s. This massive connectivity allows thousands of chips to quickly communicate with each other and access a staggering 1.77 Petabytes of shared High Bandwidth Memory (HBM), overcoming data bottlenecks for even the most demanding models

  

☑️ #211 Oct 16, 2025 🟠 opinion

Google officially starts selling TPUs to external customers

@zephyr_z9: Google officially starts selling TPUs to external customers and competes directly with Nvidia now.

⚡️

@zephyr_z9: Broadcom’s 5th customer ($10B) isn’t Apple or XAI It’s Anthropic. They won’t design a new chip. They will be buying TPUs from Broadcom. Expect Anthropic to announce a funding round from Google soon.

+ Related content:

• linkedin.com (October 2025): [Excerpt] TPU Kernel Engineer.
  As a TPU Kernel Engineer, you’ll be responsible for identifying and addressing performance issues across many different ML systems, including research, training, and inference. A significant portion of this work will involve designing and optimizing kernels for the TPU. You will also provide feedback to researchers about how model changes impact performance. Strong candidates will have a track record of solving large-scale systems problems and low-level optimization.

☑️ #210 Oct 16, 2025

Commonwealth Fusion Systems to use AI to accelerate fusion development

@demishassabis: Super excited to be collaborating with Commonwealth Fusion Systems @CFS_energy to use AI to accelerate fusion development - and move closer to a sustainable future with limitless clean energy.

⚡️

@GoogleDeepMind: We’re announcing a research collaboration with @CFS_energy, one of the world’s leading nuclear fusion companies. Together, we’re helping speed up the development of clean, safe, limitless fusion power with AI.

[2] Fusion powers the sun, but here on Earth, one approach involves controlling a super-hot, ionized gas called plasma inside a tokamak machine. To predict power generation, we need to simulate how heat, electric current and matter flow through the core of a plasma and interact with systems around it. This is where TORAX comes in.

[3] TORAX is our open-source plasma simulator allowing CFS to run millions of virtual experiments to test plans for their tokamak, SPARC. Using reinforcement learning, we’re now rapidly identifying the most efficient paths for it to generate more power than it consumes - a landmark achievement known as crossing “breakeven.”

[4] We’re also training AI agents to be expert “pilots”, discovering new ways to control the plasma in real-time, managing immense heat and maximizing energy output while staying within operating limits.

☑️ #209 Oct 16, 2025 🟥 0-shot prompting

More usage → more data → more cash

chat.openai.com (chatgpt): [Sources not verified. Do your own research]:

Here’s a detailed “flywheel” of Google’s ecosystem — showing how Search & Ads, Cloud & AI, Moonshots, and Culture each generate value, how they reinforce each other, how learning & reinvestment compound, and why the system becomes increasingly powerful (and difficult for rivals to replicate).

Search & Ads → Data + Capital

The “engine” that reaches billions of users, monetizes intent and attention. Search answers queries; ads monetize via relevance.

Massive cash flows, huge volumes of user behavior data (queries, clicks, intent), signals about what people want, trends, keywords, contexts. Also, signals about advertisers’ willingness to pay.

• Each search query, each click on an ad or content, provides signals: what people search, what they buy (or value), what content engages them.

• This data improves understanding of user intent, content relevance, ad matching.

• Ads generate large consistent cash flow. That capital can be reinvested in infrastructure and R&D.

Cloud & AI consume & refine the data

Infrastructure + intelligence layer: provides compute, storage, ML/AI platforms; deploys models; builds tools to extract insights.Algorithms, models, tools; infrastructure capabilities; enterprise sales & use cases; cross-product learning; scaling efficiencies.

• Data from Search & Ads (but also from YouTube, Android, Maps etc.) feeds into ML pipelines: cleaner data, labels, user feedback, interaction signals.

  • More accurate search / ad targeting means higher ad effectiveness → more ad spend → more revenue.

    • Cash from Ads/Search funds AI/ML R&D, infrastructure (Cloud), and Moonshots; also funds acquisitions or buying talent.

• AI models get better: better understanding of language, vision, context, personalization.

  • AI embedded into consumer products (Search, Home, Android, Maps etc.) keeps users more engaged; creates stickiness.

  • Tools/platforms refine: e.g. Search’s AI Overviews, Performance Max ads, generative AI in Workspace, etc. (see “Search & Ads innovations in the AI era”)

    • Better AI features → better user satisfaction → more usage → more data → further improved models.

• Cloud infrastructure provisioned to run these models at scale, serve products, maintain SLAs.

  • Better AI features in cloud wins enterprise customers; more usage = more feedback & more data.

  • Insights from Cloud & AI improvements feed into Search & Ads to lift performance (higher ROI), prompting more usage and more cash.

    • More enterprise adoption of Cloud/AI → more revenue → better infrastructure scale (lower cost per unit) → better margin → more investment capacity.

Moonshots (X etc.) push the frontier / provide new building blocks

Moonshots serve as incubators for future vectors.

Long‑horizon bets; radical technologies; exploration of new domains. Some will fail, but some graduate into foundational platforms.New capabilities (autonomous driving, life sciences, robotics, etc.); R&D on hard problems; breakthroughs that may feed back into core stack; domain‑specific data & architectures; future optionality.

• Moonshot R&D like Waymo, X projects, etc. work on problems that often require new sensors, new algorithms, new compute architectures, new data sources. These feed back into core capabilities.

• Even failures/non‑graduated projects often produce learnings, IP, technologies usable in more mature domains.

• They provide optionality: if a moonshot becomes viable, it can transform industries and become a major source of future data and cash (or at least defensible margins).

• Moonshots, as they mature or test components, feed into product lines or become separate business units, generating new revenue streams or at least new capabilities.

Moonshots explore new frontiers → produce new capabilities, domain data, new infrastructure, sometimes new revenue. 

Culture enables all of the above to work continuously

Culture protects against stagnation and ensures adaptability.

Norms, structure, incentives, values: curiosity, willingness to experiment & fail; long‑term thinking; integration across product lines; privacy & ethics awareness; developer talent.Speed of innovation, retention of high‑skill people, ability to pivot, trust from users and regulators, quality of products, ethical/ responsible AI practices.

• Maintaining a culture that tolerates failure, invests for long term, balances risk/reward: without that, Moonshots are hard, product innovation slows, bureaucracy kills speed.

• Ethics, privacy, responsible AI practices in culture also build trust, reducing regulatory or reputation risks that could stall the loop.

• Cross‑product team alignment (e.g., Ads, Search, Cloud, Android) helps share data, infrastructure, and models — avoiding silos.

Culture supports risk‑taking, speed, cross‑team sharing, long‑term thinking, ethical guardrails. 

☑️ #208 Oct 16, 2025 🟠 opinion

Value of the ad placed in AI mode will be much higher for advertisers, as users provide more insights into what they are looking for

@RihardJarc: Some golden nuggets from an interview with a former $GOOGL employee on TPUs and the future of Search:

1. According to him, $NVDA is specifically bought to satisfy customer demand (on the cloud). $GOOGL doesn’t use $NVDA GPUs for production use cases on their own products like Gemini. Every first-party service that $GOOGL has is powered by TPUs.

2. $GOOGL uses TPUs because they are cheaper on a performance basis, you don’t have to rely on $NVDA, and because the entire $GOOGL stack is optimized for TPUs all the way to the top.

3. According to him, $GOOGL is investing heavily to build out more TPUs and to increase capacity by third-party data providers so that they can actually run their cloud.

4. When it comes to Search, two years ago he would be pessimistic, but he has now been positively impressed by $GOOGL in the last 9 months and how well $GOOGL has integrated AI into their search UI. Not just speed but also »the gutsiness of it, because as you can imagine, they’re eating up a lot of ad space«. They are seeing a lot of out-of-clicks, paid clicks not growing as fast, whereas CPCs are going up much faster. Inventories are being squeezed out due to much less ad space, and as a result, prices are also rising.

5. He does see ads as a business model in the new AI search experience. He believes the value of the ad placed in AI mode will be much higher for advertisers, as users provide more insights into what they are looking for.

found on @AlphaSenseInc

☑️ #207 Oct 13, 2025

A comprehensive snapshot of GOOGLE’s (Alphabet) current position and recent developments

perplexity.ai: [Sources not verified. Do your own research]

Financial Performance and Market Standing

• Alphabet’s stock (GOOG) is currently trading near $244.64 per share, with a market capitalization standing at approximately $2.96 trillion.

• The company has seen steady revenue growth, reporting $96.43 billion in revenue for the recent quarter, exceeding estimates and reflecting a year-on-year growth rate of 13.8%.stockstory​

• Core advertising (Google Search and YouTube) remains the primary driver, but Google Cloud continues as a fast-growing business, with its five-year annualized growth for Cloud at 35.5% and recent two-year growth at 28.8%.stockstory​

• Alphabet boasts a low dividend yield (0.34%), reflecting its emphasis on reinvestment over payouts.

Recent Developments

• In this October 2025, Alphabet announced a $15 billion investment to build a major artificial intelligence data hub in India, reflecting its continued leadership and expansion in AI infrastructure worldwide.qna​

• The stock recently surged following a favorable legal ruling, further strengthening investor confidence and contributing to positive market sentiment. On October 9, the stock closed at $242.21 per share with a $2.925 trillion market cap.finance.yahoo​

• Alphabet is set to report its Q3 2025 financial results in late October.abc​

Competitive Landscape and Strategic Priorities

• Google Search retains more than 80% global market share, but faces competition from Microsoft’s Bing (boosted by OpenAI’s ChatGPT integration) and Meta’s AI-driven advertising advancements.stockstory​

• YouTube continues as a major asset, providing significant advertising revenue and engagement.finance.yahoo+1​

• Google Cloud, while trailing Amazon’s AWS by market share, is a key focus area for growth and innovation, supporting enterprise digital transformation.stockstory​

Industry Challenges

• Alphabet faces regulatory scrutiny in the U.S. and abroad, as large tech companies undergo increased antitrust examination and data privacy regulation.investing​

• The rise of generative AI and competitive shifts in digital advertising algorithms present both disruption risks and opportunities for Alphabet to leverage its uniquely vast data resources.investors+2​

Alphabet stands as a dominant, resilient tech conglomerate—diversifying rapidly, investing heavily in the future of AI, and maintaining strong financial health amid evolving regulatory and competitive pressures.qna+3​

☑️ #206 Oct 7, 2025

When will the Transformer earn its Nobel Prize?

@chr1sa: Google is now the #2 company in the history of Nobel Prizes, passing IBM

+ Related content:

• @ComputerBoardEE: Interesting who gets to claim credit. The original work was done at Berkeley, then JM went to NIST, and then spent a decade building a research group at UCSB that google hired as their quantum computer division in SB.

  ​

• @Google: Congratulations to Michel Devoret, Google Quantum AI’s Chief Scientist of Quantum Hardware, who was awarded the 2025 Nobel Prize in Physics today. Google now has five Nobel laureates among our ranks, including three prizes in the past two years.

  

• @Google: The five Nobel laureates include current Googlers: -Demis Hassabis (Chemistry, 2024) -John Jumper (Chemistry, 2024) -Michel Devoret (Physics, 2025) And Google alums: -Geoff Hinton (Physics, 2024) -John Martinis (Physics, 2025) Congratulations!

  ​

  • @Yuchenj_UW: Google has won 3 Nobel Prizes in just 2 years: >2024: Demis Hassabis (AlphaFold), Geoff Hinton (AI) >2025: Michel & Jone (Quantum Computing) No other tech company matches Google’s long-term research taste and ambition. Not Apple, Microsoft, or Amazon. The real question: when will the Transformer earn its Nobel Prize?

☑️ #205 Oct 4, 2025

The Berkshire of Tech

@sahilypatel: google is the berkshire of tech

• owns 14% of anthropic

• owns 8% of spacex

• acquired double click for $3.1b (backbone of google ads empire)

• acquired youtube for $1.65b in 2006 (worth hundreds of billions today)

• acquired android in 2005 for $50m; now the world’s most widely used os.

• acquired deepmind in 2014 for $500m, huge ai lab with almost sota models and great research team.

• runs x (the moonshot factory), incubator of radical ideas (waymo, loon, wing drones, etc.)

• through capitalg has invested in stripe, uipath, airbnb, lyft, databricks, crowdstrike, duolingo, robinhood and more

⚡️

@scottbelsky: not to mention Waze, and now Wiz (security juggernaut)

• @sahilypatel: ah, missed those. waze and wiz def belong on the list

⚡️

@connebs: > deepmind > *almost* SOTA There is more SOTA work coming out of deepmind than any other lab on earth

• @sahilypatel: yeah my bad. i meant that pretty much every model they launch is sota

+ Related content:

• @FIREDUpWealth: Fun fact: Alphabet owns 14% of Anthropic and 8% of SpaceX. So if you own $GOOGL or $GOOG stock, you own a part of each

  

☑️ #204 Oct 3, 2025

Gradient Ventures

Congrats to my cousin @darian314 for leading Gradient Ventures to successfully spin out of @Google.

Corporate venture funds rarely succeed, and even more rare is to succeed enough to spin out.

Alphabet (Google) founded Gradient Ventures in 2017 for technical mentorship for early-stage startups focused on AI.

☑️ #203 Sep 5, 2025

EU Competition

@EU_Competition: Today the EU Commission has fined Google €2.95 billion over abusive practices in the online advertising technology industry, AdTech. Google distorted competition by favouring its own services, harming rivals, advertisers and publishers. More info

https://europa.eu/!ytrqyV

☑️ #202 Sep 3, 2025

TPUs and puzzling questions

@BenBajarin: The idea of Google looking to shop TPUs is actually not a good sign. They still can't get adoption on the CSP portion of GCP. So did they overbuild and are looking to recoup those? Are they prioritizing GPUs and not TPUs internally for more workloads so they have excess inventory?

Regardless, if GCP is having trouble getting software devs to lift and shift for TPU, why would Neoclouds have any more success?

Lot's of puzzling questions.

⚡️

Are they prioritizing GPUs and not TPUs internally for more workloads so they have excess inventory? 

phind.com (Phind-70B): [Sources not verified. Do your own research]: No, Google is not prioritizing GPUs over TPUs internally - quite the opposite. Recent evidence shows Google is heavily investing in TPUs and maintaining a clear strategic separation between their internal TPU usage and external GPU offerings.

Current Strategy

• Google runs TPUs exclusively for internal workloads while offering Nvidia GPUs to cloud computing customers

• They maintain the largest installed base of AI accelerators across the industry

• TPU shipments reached 2 million units last year alone

Strategic Investment Evidence

1. Market Position - Google has become the third-largest designer of datacenter processors, behind only Nvidia and Intel

   1. Their TPU business accelerated significantly with each new generation, especially after TPU v4 (2021)

2. Infrastructure Growth - The company is expanding its TPU infrastructure significantly

   1. They're preparing for "explosive growth" of large language models like Gemini

   2. New TPU generations (V5e and V5p) are being deployed more widely than previous ones

Technical Optimization Focus

Google continues to optimize TPUs specifically for their internal workloads:

• They've achieved a 3x improvement in carbon efficiency across TPU generations

• Operational electricity emissions comprise over 70% of TPU lifetime emissions

• The company is actively working to improve manufacturing sustainability

Rather than having excess inventory, Google is strategically expanding their TPU infrastructure while maintaining a clear separation between their internal TPU usage and external GPU offerings for cloud customers. This dual-strategy approach aligns with their position as a leading provider of both custom silicon and cloud computing services.

+ Related content:

• arxiv.org (2/10/25): [Abstract] Machine Learning Fleet Efficiency: Analyzing and Optimizing Large-Scale Google TPU Systems with ML Productivity Goodput.

  Recent years have seen the emergence of machine learning (ML) workloads deployed in warehouse-scale computing (WSC) settings, also known as ML fleets. As the computational demands placed on ML fleets have increased due to the rise of large models and growing demand for ML applications, it has become increasingly critical to measure and improve the efficiency of such systems. However, there is not yet an established methodology to characterize ML fleet performance and identify potential performance optimizations accordingly. This paper presents a large-scale analysis of an ML fleet based on Google’s TPUs, introducing a framework to capture fleet-wide efficiency, systematically evaluate performance characteristics, and identify optimization strategies for the fleet. We begin by defining an ML fleet, outlining its components, and analyzing an example Google ML fleet in production comprising thousands of accelerators running diverse workloads. Our study reveals several critical insights: first, ML fleets extend beyond the hardware layer, with model, data, framework, compiler, and scheduling layers significantly impacting performance; second, the heterogeneous nature of ML fleets poses challenges in characterizing individual workload performance; and third, traditional utilization-based metrics prove insufficient for ML fleet characterization. To address these challenges, we present the “ML Productivity Goodput” (MPG) metric to measure ML fleet efficiency. We show how to leverage this metric to characterize the fleet across the ML system stack. We also present methods to identify and optimize performance bottlenecks using MPG, providing strategies for managing warehouse-scale ML systems in general. Lastly, we demonstrate quantitative evaluations from applying these methods to a real ML fleet for internal-facing Google TPU workloads, where we observed tangible improvements.

• cloud.google.com: Accelerate AI development with Google Cloud TPUs

  • cloud.google.com (2/6/25): Designing sustainable AI: A deep dive into TPU efficiency and lifecycle emissions.

• Next’24

☑️ #201 Sep 3, 2025

GOOGL 5%+ Non-Earnings Related Gap Openings

@bespokeinvest: Since its IPO in 2004, $GOOGL has gapped up at least 5% on 31 different days. Only nine were not in reaction to earnings.

+ Related content:

• Michael Spencer (): Breaking News: Google gets to keep Chrome but is barred from exclusive deals.

  Pretty much the expected ruling but still significant since browser-based AI tools can allow AI agents to be more autonomous.

  So essentially, Google was found to hold an illegal monopoly in its core market of internet search. But still gets to keep the majority of its Monopoly power. There's about 3 billion users of Chrome in the world.

  Google stock should be up 5 to 8% tomorrow and popped after hours. Google's really starting to lead in generative AI and a generative AI product infused ecosystem.

  This kind of slap on the wrist ruling is also good for Meta and its antitrust case. The US does not appear serious around regulating its monopolies. The biggest loser is Apple that won't have that 20 billion contract for search.

  As you might know, Google pays Apple billions of dollars per year to be the default search engine on iPhones.

☑️ #200 Sep 3, 2025

Google will not be forced to sell Chrome but has to share search data

The Lunduke Journal of Technology: Google Dodges All Remedies in Search Engine Monopoly Case. Judge Mehta rules that Google will not need to sell Chrome. Or Android. And will be allowed to continue providing 80%+ of Mozilla's revenue. In fact, no real punishment at all.

The Lunduke Journal of Technology

Google Dodges All Remedies in Search Engine Monopoly Case

Judge Amit Priyavadan Mehta, a 2014 appointee of President Barack Hussein Obama II, has issued his long anticipated, final ruling in the remedy phase in the case of USA v. Google…

Read more

a year ago · 10 likes · 1 comment · Bryan Lunduke

The Lunduke Journal of Technology

The Lunduke Journal is the last bastion of truly independent Tech Journalism. Ad Free, Big Tech Free, Audience Supported.

By Bryan Lunduke

+ Related content:

• @WebDesignMuseum (9/2/25): On September 2, 2008, Google released the beta version of Google Chrome’s freeware browser for Windows. In December 2008, the first stable version of Google Chrome 1.0 was released.

  #InternetHistory

  

• @ddayen (9/2/25): Judge Mehta on the proposed Chrome divestiture as a remedy in Google's monopolization case:

  1-Chrome definitely contributes to Google's search monopoly

  2-It's not even as radical as the Microsoft proposed remedy

  3-It would hardly break up the company 4-But let's not do it

  @ddayen: I seriously doubt Google would appeal this. DOJ should as it doesn't touch Google's monopoly in search, for which it was found guilty. But there's good reason to doubt Big Tech errand boy Trump won't let that appeal go through

  @ddayen: DOJ: Google paid distributors for default placement, entrenching its monopoly

  Judge (verdict): Yes, illegal conduct

  DOJ: As a remedy, you must prohibit paying distributors for default placement

  Judge (remedy): "Google will be permitted to pay distributors for default placement."

☑️ #199 Aug 27, 2025

Ironwood Architecture (TPUv7)

@JeffDean: My Google colleagues Norm Jouppi & Sridhar Lakshmanamurthy gave a talk today at Hot Chips on TPUv7 ("Ironwood"). The TPUv7 system offers 9216 chips / pod (42.5 exaflops of fp8), but we can scale across many of these pods to provide multiple zettaflops.

⚡️

@SemiAnalysis_: Google presents for the first time ever their TPUv7 block diagram at hot chips conference. TPUv7 (formerly known as TPUv6p, internally called ghostfish) has 8 stacks of HBM3e memory, 4 medium size systolic arrays and be connected in a 3D torus with a scale up world size of up to 9216. Note that even with OCS, the blast radius of 3D torus is quite huge and a failure of 1 chip takes out a whole cube of 64 TPU chips. Previously without OCS, a failure of 1 chip in an 3d torus would take the whole cluster such as in TPUv3

+ Related content:

• @bubblebabyboi (8/26/25): Nvidia will deny this lol. Of the top ten models 2-3 are not trained on Nvidia GPUs… for inference it’s probably worst.

  • @jxnlco: interesting

    

  ​

• blog.google (4/9/25): [Excerpt] Ironwood: The first Google TPU for the age of inference.

  Ironwood is designed to gracefully manage the complex computation and communication demands of "thinking models," which encompass Large Language Models (LLMs), Mixture of Experts (MoEs) and advanced reasoning tasks. These models require massive parallel processing and efficient memory access. In particular, Ironwood is designed to minimize data movement and latency on chip while carrying out massive tensor manipulations. At the frontier, the computation demands of thinking models extend well beyond the capacity of any single chip. We designed Ironwood TPUs with a low-latency, high bandwidth ICI network to support coordinated, synchronous communication at full TPU pod scale.

  For Google Cloud customers, Ironwood comes in two sizes based on AI workload demands: a 256 chip configuration and a 9,216 chip configuration.

  

☑️ #198 Aug 20, 2025

Pixel 10

@madebygoogle: Live From NYC, It’s #MadeByGoogle ‘25

+ blog.google/products/pixel/

☑️ #197 Aug 3-4, 2025

Annual Net Income Leader: $116B

@StockSavvyShay (8/3/25): $GOOGL IS NOW THE MOST PROFITABLE COMPANY IN THE WORLD WITH $116B IN ANNUAL NET INCOME.

⚡️

@IrnestKaplan (8/4/25): $GOOG Net Income Leader It has come up from third place in 2023. Understandably, $NVDA has rocketed up from last to 4th place over the past 2.5 years. Lets see how it looks a few years from now..

☑️ #196 Aug 1, 2025

Google AI model mines trillions of images to create maps of Earth ‘at any place and time’

nature.com: [Excerpt] The system will save time spent on processing satellite data, researchers say — but they hope the tech firm will share more about how it works.

The model, called AlphaEarth Foundations, maps the world “at any place and time”, Christopher Brown, a research engineer at Google DeepMind in New York City, said at a press briefing on 28 July. Scientists might use it to identify the best places for clean-energy projects, study the ecological impacts of climate change, track deforestation and more.

Alongside the model, which was released yesterday, Google and its DeepMind AI unit have posted a paper, ahead of peer review, to the arXiv preprint server¹. The manuscript describes the creation of the model, its initial testing and baseline data sets covering Earth during the years 2017–24. It says that AlphaEarth beats similar AI models in terms of both accuracy and data density, and that the model saves users computing time and reduces energy use.

+ Continue reading

+ Related content:

• deepmind.google (7/30/25): [Excerpt] AlphaEarth Foundations helps map our planet in unprecedented detail.

  To ensure AlphaEarth Foundations was ready for real-world use, we rigorously tested its performance. When compared against both traditional methods and other AI mapping systems, AlphaEarth Foundations was consistently the most accurate. It excelled at a wide range of tasks over different time periods, including identifying land use and estimating surface properties. Crucially, it achieved this in scenarios when label data was scarce. On average, AlphaEarth Foundations had a 24% lower error rate than the models we tested, demonstrating its superior learning efficiency. Learn more in our paper.

  

​

💲 EARNINGS ANNOUNCEMENT: Jul 23,2025

Alphabet Announces Second Quarter 2025 Results (pdf)

Webcast & Slides (pdf)

⇢ Some reactions:

• Global Equity Briefing (update; 7/24/25): Google Q2 Earnings Update! Growth accelerates, and Search is alive and kicking!

  Simply put, the narrative that Google’s Search business is dead is dead!

  Global Equity Briefing

  Google Q2 Earnings Update!

  Read more

  a year ago · 13 likes · 2 comments · Ray Myers

• @pauldejong_: Google's underlying earnings power has expanded at an even faster rate than its market price. Earnings growth is the true powerhouse behind Google. $GOOG

  

☑️ #195 Jul 1, 2025

"Contrary to Google's contentions, the district court specifically noted trial evidence showing the anticompetitive nature of anti-steering restrictions.' Those anticompetitive effects, if the restrictions were not enjoined, would continue to harm competition in the defined markets of Android in-app billing and Android app distribution, in which Epic is undisputedly a player,"

ca9.uscourts.gov/opinions: EPIC GAMES, INC. V. GOOGLE LLC, ET AL. Case no. 24-6256 (pdf)

☑️ #194 Jul 15, 2025

Introducing Rivian Navigation with Google Maps

stories.rivian.com: [Excerpt] Providing the best of both worlds for our drivers.

What’s new from Google Maps

Rivian drivers will benefit from fresh traffic and place information from Google Maps. This means access to reliable routes directly in their Rivian vehicle. Updated features include:

Fresh routes and traffic: No matter the destination, fresh traffic data will help you find the fastest route, keep your ETA accurate along the way, and even reroute you if a quicker option becomes available during your drive.

Real-time information about road disruptions: Whether it's a lane closure for construction or an accident, stay informed about the road ahead with real-time disruption reports from the Google Maps community.

Tappable points of interest (POIs): Rivian has already been using Google Maps for search to bring you detailed information on places, like hours, photos, ratings, and place summaries. Now, you can tap on places of interest right on the map and learn more about hours, reviews, and contact information with just a tap. This builds on our existing place search experience that already uses Google Maps’ information about 250 million places around the world to deliver helpful information.

Satellite imagery: With satellite imagery from Google Maps, you can easily orient yourself in the city or scout the perfect location for your next adventure.

+ Continue reading

☑️ #193 Jun 24, 2025

Repurchase of Common Stocks

@KoyfinCharts: After 7 years of buybacks and a 13% reduction in share count, Alphabet has returned its share count to levels last seen in 2006.

+ Related content:

• app.koyfin.com: Buyback Yield %

  

  • koyfin.com > Glossary: [Excerpt] Buyback Yield % is the repurchase of outstanding shares over the existing market cap of a company. It is calculated using the formula:
    
    Buyback Yield % = ((Issuance of Common Stock + Repurchase of Common Stock) / Latest Market Cap)​

☑️ #192 Jun 19, 2025

Search’s rapid decline

@qcapital2020: Google to zero.

+ Related content:

• similarweb.com > google.com

  • search.google.com

☑️ #191 Jun 19, 2025

Google is using some of the 20B videos on YouTube to train its AI tool

cnbc.com: [Excerpt] Creators say they didn’t know Google uses YouTube to train AI.

• Google is using its expansive library of YouTube videos to train its AI models, including Gemini and the Veo 3 video and audio generator, CNBC has learned.

• The company confirmed to CNBC that it relies on its vault of YouTube videos to train its AI models, but Google said it only uses a subset of its videos for the training.

• Such use of YouTube videos has the potential to lead to an intellectual property crisis for creators and media companies, experts said.

Google is using its expansive library of YouTube videos to train its artificial intelligence models, including Gemini and the Veo 3 video and audio generator, CNBC has learned.

The tech company is turning to its catalog of 20 billion YouTube videos to train these new-age AI tools, according to a person who was not authorized to speak publicly about the matter. Google confirmed to CNBC that it relies on its vault of YouTube videos to train its AI models, but the company said it only uses a subset of its videos for the training and that it honors specific agreements with creators and media companies.

+ Continue reading

☑️ #190 Jun 12, 2025

@The_AI_Investor: Google is the only fully vertically integrated player.

⚡️

@ArtificialAnlys: Google is firing on all cylinders across AI - Gemini 2.5 Pro is equal #2 in intelligence, Veo 3 and Imagen 4 are amongst the leaders in media generation, and with TPUs they're the only vertically integrated player

Google is now equal #2 Artificial Analysis Intelligence Index with the recent release of the Gemini 2.5 Pro (June 2025) model, rivaling others including OpenAI, DeepSeek and Grok

Google Veo 3 now ranks second in the Artificial Analysis Video Arena Leaderboard only behind ByteDance’s new Seedance 1.0 model

Google Imagen 4 now occupies 2 out of the top 5 positions on the Artificial Analysis Image Arena Leaderboard

Google has a full stack AI offering with offerings across the application layer, models, cloud inference and hardware TPUs)

@ArtificialAnlys: Google has consistently been shipping intelligence increases in its Gemini Pro series

@ArtificialAnlys: Google Veo 3 now occupies second place in the Artificial Analysis Video Arena, after originally debuting in the first place. Still a significant leap over Google Veo 2!

@ArtificialAnlys: Google Imagen 4 Ultra Experimental and Imagen 4 Preview now rank 4th and 5th respectively in the Artificial Analysis Image Arena.

@ArtificialAnlys: Further analysis in our quarterly State of AI report (highlights report freely accessible on Artificial Analysis)

Language models analysis: https://artificialanalysis.ai/models

Image arena: https://artificialanalysis.ai/text-to-image/arena

Video Arena: https://artificialanalysis.ai/text-to-video/arena

☑️ #189 May 12, 2025

Helping startups build what’s next with the AI Futures Fund

blog.google: [Excerpt] Early participants.

Over the past several months, the AI Futures Fund has been working with a select group of startups across various industries:

+ Continue reading

+ Related content:

• labs.google: AI Futures Fund. Working with startups to build what’s next in AI.

  • Early access to select AI models

  • Google expertise

  • Equity funding

  • Cloud credits and support

☑️ #188 May 10, 2025

GOOGL: Is this it for Search?

@rwang07: BERNSTEIN: Is this it for Search ( $GOOGL)?

Over the past two weeks, data points highlighting the slowest paid click growth in Google’s history combined with the first month ever of declining web search volumes on Apple devices validate investors’ feeling – it’s not just be them changing their search behavior.

1. Growing pies and shrinking slices. There are two truths: (1) the “search” pie is growing, and (2) Google’s share is shrinking. The question that matters most is whether Google’s slice of the larger pie is still bigger than what their slice would have been pre Gen AI.

2. Is a growing pie commercial? Gen AI is adding queries we never asked Google before – create an image, draft an email response, and summarize topics. Most of these appear difficult to monetize, though Gen AI should also migrate more commercial queries online – if we can research better, we should convert at a higher rate.

3. Shrinking slices come in threes. AI user number updates from ChatGPT and MetaAI creep towards 1B, lowest ever paid click growth on Google, and Apple seeing first month decline in search volume. Combined, we estimate Google’s search share is closer to 65-70% vs. the 90% we often hear.

4. Google Search is now social. Investors scrutinize exposure growth to evaluate terminal risk/potential of social media rails – we did this with Meta and Snapchat in the face of TikTok download share gains. We should do the same with Google – Google is now firmly in this camp with revenue growth now viewed as a lagging indicator.

   

⚡️

@MonkEchevarria: Man, that Bernstein report on Google’s slipping search grip is a real eye-opener. I’m cautiously optimistic about GOOG, though. Sure, market share is down, and AI rivals are hungry, but their profitability is still a beast, ranking high against peers. Even with a pricey valuation dragging its value score, their cash flow machine might just keep ‘em ahead. Gotta watch that shrinking pie slice, tho; it’s a wild tech jungle out there.

☑️ #187 May 7, 2025

The $164 zone

@VolumeLeaders: $GOOG / $GOOGL - It would be disingenuous for me to say this bad news and subsequent drop was knowable in advance. But what I can demonstrate is that the level from which the decline began is a very strong s/r level that goes all the way back to last April. It's its 4th largest level since inception. VL provides this context by painting levels in which elevated institutional activity has occurred in the past.

Sept 2024 OPEX (and to a lesser degree April 2024 EOM) came in quite large and the level in which that inventory traded held as support on multiple tests from above going into EOY. When the 2025 decline began, it served as temp support before ultimately breaking lower.

Upon revisiting that level from below, it now appears to have held once more, just in the opposite direction.

The news itself wasn't foreseeable, but if news was going to come (positive or negative), this location is a pretty good candidate for where we'd see it. In other words, the $164 zone was a good place to reduce risk or trail stops if long given how much activity has transpired here in the past. Should that level be recaptured in the future, it may well serve as strong support once again. Time will tell.

volumeleaders.com/register

⚡️

@VolumeLeaders: $GOOG - Here's a closer look at the action since March. Several interactions with this level.

+ Related content:

• sherwood.news: [Excerpt]Google tanks, dragging market lower, after Apple says it’s considering adding AI search in its browser.

  • APPLE’S CUE: SEARCHES IN BROWSER FELL FOR FIRST TIME IN APRIL

  • APPLE’S CUE: APPLE EXPLORING ADDING AI SEARCH TO ITS BROWSER

  • APPLE’S CUE: BELIEVES PERPLEXITY, ANTHROPIC COULD BE OPTIONS

• justice.gov/atr > U.S. and Plaintiff States v. Google LLC [2020] (Remedies Hearing) (5/2/25): [Google Is a Monopolist in Ad Tech: May 2 hearing date to discuss potential remedies in the antitrust case] United States District Court for the Eastern District of Virginia Leonie Brinkema set Alphabet Inc.'s Google's newest antitrust trial related to its advertisement policy for September 22.

☑️ #186 May 1, 2025

Over 55M miles of fully autonomous driving data

@dmitri_dolgov: Over 55M miles of fully autonomous driving data shows the @Waymo Driver is making roads safer for everyone, especially for those most at risk – pedestrians, cyclists, and other vulnerable road users. Strong positive impact already and growing exponentially. waymo.com/blog/2025/05/waymo-making-streets-safer-for-vru

+ Related content:

• waymo.com (5/1/25: [Excerpt] New Study: Waymo is reducing serious crashes and making streets safer for those most at risk.

  The path to Vision Zero requires reducing severe crashes and improving the safety of those most at risk. Our latest research paper shows that the Waymo Driver is making significant strides in both areas. By reducing the most dangerous crashes and providing better protection for pedestrians, cyclists, and other vulnerable road users, Waymo is making streets safer in cities where it operates.

  The paper, accepted to be published in the Traffic Injury Prevention Journal, expands on our Safety Impact Hub research, providing a deeper analysis of Waymo’s performance across 11 different crash types compared to human drivers. It also offers new insights into Waymo’s positive impact on serious injury crash rates.

​

💲 EARNINGS ANNOUNCEMENT: Apr 24,2025

Alphabet Announces First Quarter 2025 Results (pdf)

Webcast

⇢ Some reactions:

• Bill Ackman - Pershing Square Capital Management, L.P. (via Uncle Stock Notes): [Excerpt] Revenues grew at a robust 15% rate led by strength in its core advertising business and rapid deployment of Al innovations. In Search, the company has successfully rolled out Al-powered summary responses, called "Al Overviews". Early engagement with Al Overviews has been highly encouraging with consumers searching more frequently, with more detailed queries, and clicking through at higher rates. Google has been able to address two key investor concerns with introducing Al in Search around cost-to-serve and ad monetization. The company has reduced the compute cost per Al query by more than 90% over the last 18 months, a step-function cost reduction which has enabled Google to roll out Al Overviews broadly to more than a billion users. The company also recently introduced ads on Al Overviews and has explained that these ads are already monetizing at the same rate as regular Search ads. Over time, we expect more context-rich Al responses to drive even higher conversion rates and efficacy on ads.

  Beyond Search, Google's Al capabilities continue to differentiate and strengthen the company's other business lines. Al-powered creative tools have propelled YouTube to be a leading destination for content creators and driven its success in the short-form video format and on Connected TVs, where YouTube is the number one streaming platform in the U.S. by watch time according to Nielsen. Google's Cloud segment, which is now a nearly $50 billion run-rate business, grew 31% in 2024 driven by demand for the company's enterprise Al offerings. Most notably, the segment's impressive growth rate was potentially capped by near-term capacity constraints, which the company is seeking to alleviate with its expanded $75 billion capital expenditure plan in 2025. With increasing scale, the Cloud segment has inflected from breakeven profitability in 2023 to 17.5% profit margins in the most recent quarter, with line of sight to achieving the greater than 30% profit margins enjoyed by peers.

  More broadly, profitability across Alphabet is improving at a rapid clip. Consolidated profit grew 29% and profit margins expanded 370 basis points in 2024 (excluding one-time severance and restructuring charges) as cost control initiatives and headcount leverage are bearing fruit. Over the last two years, the company's total headcount has decreased 4% while revenue has grown 27%. In addition to driving revenue growth opportunities, Al integration is also enabling substantial automation gains in Google's internal workflows and operations. The company recently shared that more than a quarter of all new code is generated by Al and then reviewed and accepted by engineers. We expect these automation efficiencies to continue driving margin expansion over the long-term. To that end, we were encouraged to hear new CFO, Anat Ashkenazi, commit to building on the company's efficiency initiatives and evaluate where they might be able to accelerate their work.

  Despite tremendous business momentum, Alphabet still trades at around 18x forward earnings, a highly discounted valuation for a business of its quality and growth prospects. As the company executes on its Al potential and margin expansion opportunity to drive rapid earnings growth, we expect its share price will increase to better reflect its strong underlying business fundamentals.

  • marketwatch.com (5/25/23): [Excerpt] ‘An unmatched business model’: Why Bill Ackman’s Pershing Square spent $1 billion on Alphabet stock.

    Pershing disclosed earlier in May that it amassed a new position in Alphabet during the first quarter that amounted to 10.3 million shares across the Class A and Class C classes. That position was worth upwards of $1 billion as of the end of the quarter, though Ackman said Wednesday that Pershing’s average cost was about $94 to $94.50 a share.

​

☑️ #185 Apr 23, 2025 🟠 opinion

District Court Rules Google Is a Monopolist in Ad Tech

stblaw.com: [Excerpt] On April 17, 2025, U.S. District Judge Leonie Brinkema of the U.S. District Court for the Eastern District of Virginia issued a decision in the liability phase of the Google ad tech monopolization case brought by the Department of Justice (“DOJ”) and 17 states (collectively “Plaintiffs”). In an opinion that chronicled the history and evolution of web display advertising technology, the court found Google liable for monopolizing both the market for open-web display ad exchanges, which conduct the real-time auctions to place display advertisements on websites, and the market for ad servers used by website publishers, which manage the sale and display of advertisements on the publisher’s website. It also found Google liable for tying its products in the two markets together. The case will now move on to the remedy phase, with the parties due to submit their proposed schedules this week. Notably, the prayer for relief in the DOJ’s complaint indicates they planned to seek the divestiture of both DoubleClick and AdX.

+ Continue reading

+ Related content:

• vaed.uscourts.gov: Court Locations.

• justice.gov (4/17/25): [Excerpt] Department of Justice Prevails in Landmark Antitrust Case Against Google.

  Today the Antitrust Division of the Department of Justice prevailed in its second monopolization case against Google. In United States et al. v. Google, the U.S. District Court for the Eastern District of Virginia held that Google violated antitrust law by monopolizing open-web digital advertising markets. According to the Court, Google “harmed Google’s publishing customers, the competitive process, and, ultimately, consumers of information on the open web.”

  • https://www.lexology.com: Click here to read the U.S. Federal District Court for the Eastern District of Virginia Court’s decision in U.S. v. Google.

  • justice.gov/atr > U.S. and Plaintiff States v. Google LLC [2020] (Remedies Hearing) [Excerpt] The U.S. filed suit to stop Google from protecting its search and advertising monopolies through a series of exclusionary agreements designed to lock up the primary avenues through which users access search engines. Trial began on September 12, 2023. Trial exhibits and documents are available on the case page.

• usvgoogleads.com: [Excerpt] The antitrust trial of the United States v. Google, LLC. Google is accused of illegally monopolizing the digital advertising industry where they hold a mob-like control over who gets to earn revenue from ads, often depriving real journalism while allowing disinformation to profit.

  

☑️ #184 Apr 23, 2025

Perplexity has been asked by both the DOJ and Google to provide input on this case

perplexity.ai: [Excerpt] Choice is the Remedy.

Last August, the US Government ruled that Google illegally exercised monopoly powers by controlling which apps are on your phone, and worse, which apps are NOT on your phone.

Google controlled this through elaborate contracts with phone makers and carriers. The terms of the contracts threatened OEMs and carriers with wide ranging penalties for not satisfying vague or arbitrary reviews by Google.

This week, Google will be back in court with the Department of Justice to determine what the remedy for this should be.

Perplexity has been asked by both the DOJ and Google to provide input on this case. That’s rare. It tells you this moment matters—because what happens here affects the future of innovation, not just in search, but across the digital economy.

+ Continue reading

☑️ #183 Apr 9, 2025

Ironwood: The first Google TPU for the age of inference

news.samsung.com: [Excerpt] Ironwood is our most powerful, capable and energy efficient TPU yet, designed to power thinking, inferential AI models at scale.

Ironwood is built to support this next phase of generative AI and its tremendous computational and communication requirements. It scales up to 9,216 liquid cooled chips linked with breakthrough Inter-Chip Interconnect (ICI) networking spanning nearly 10 MW. It is one of several new components of Google Cloud AI Hypercomputer architecture, which optimizes hardware and software together for the most demanding AI workloads. With Ironwood, developers can also leverage Google’s own Pathways software stack to reliably and easily harness the combined computing power of tens of thousands of Ironwood TPUs.

+ Continue reading

+ Related content:

• blog.google: Google Cloud Next 25

• cloud.google.com: [Excerpt] Cloud Tensor Processing Units (TPUs). Accelerate AI development with Google Cloud TPUs.

  Cloud TPUs optimize performance and cost for all AI workloads, from training to inference. Using world-class data center infrastructure, TPUs offer high reliability, availability, and security.

  • Announcing Ironwood, our most powerful, capable and energy efficient TPU yet, designed to power thinking, inferential Al models at scale.

☑️ #182 Apr 9, 2025

“Hey Ballie”

news.samsung.com: [Excerpt] Samsung and Google Cloud Expand Partnership, Bring Gemini to Ballie, a Home AI Companion Robot by Samsung.

Gemini on Google Cloud will enable Ballie to bring personalized interactions and proactive home assistance to users; available in United States and Korea this Summer.

+ Continue reading

+ Related content:

• samsung.com: Ballie is more than a home robot, it's your buddy.

  

☑️ #181 Mar 29, 2025 🟠 opinion

The Mediocrity of Modern Google

om.co: [Excerpt] The signs are clear for those willing to see them. Like Microsoft in the late ’90s and IBM before that, Google is proving ill-equipped to respond to fundamental shifts in user behavior. The company remains trapped in its “10 blue links” jail, even as the world moves toward direct, conversational AI interfaces. Despite helping pioneer the technologies that enabled this transition, Google seems unable to break free from its advertising-driven business model.

+ Continue reading

☑️ #180 Mar 25, 2025

CME Group Will Introduce Tokenization Technology to Enhance Capital Market Efficiency Using Google Cloud's New Universal Ledger

investor.cmegroup.com: [Excerpt] CHICAGO and SUNNYVALE, Calif., March 25, 2025 /PRNewswire/ -- Further expanding their transformational partnership, CME Group and Google Cloud today announced that they are piloting solutions for seamless and secure wholesale payments and tokenization of assets. CME Group has successfully completed the first phase of integration and testing of Google Cloud Universal Ledger (GCUL), a novel distributed ledger from Google Cloud.

+ Continue reading

☑️ #179 Mar 18, 2025

Google DeepMind and NVIDIA are working to build trust in generative AI through content transparency

nvidianews.nvidia.com: [Excerpts] Developing Responsible AI and Open Models.

Google DeepMind and NVIDIA are working to build trust in generative AI through content transparency.

• To power research and AI production efforts for its customers, Alphabet’s Google Cloud will be among the first to adopt the NVIDIA GB300 NVL72 rack-scale solution and NVIDIA RTX PRO™ 6000 Blackwell Server Edition GPU, also announced today at GTC.

• NVIDIA will be the first industry partner to adopt SynthID, a Google DeepMind AI technology that embeds digital watermarks directly into AI-generated images, audio, text or video.

• Intrinsic is an Alphabet company focused on making intelligently adaptive AI for robotics usable and valuable for manufacturers across industries. Today, the majority of the world’s installed industrial robots are manually programmed, with every movement hard-coded in a complex, expensive process.

  Partnering with NVIDIA, the teams have built deeper and more intuitive developer workflows for Intrinsic Flowstate to support NVIDIA Isaac Manipulator foundation models for a universal robot grasping capability. Using foundation models for robotics will significantly reduce application development time and improve flexibility, with AI that can adapt effortlessly. At GTC, Intrinsic will also share an early OpenUSD framework streaming connection between Intrinsic Flowstate and NVIDIA Omniverse — enabling real-time visualization of robot workcells across platforms.

+ Continue reading

☑️ #178 Mar 11, 2025

Silicon Valley riders: A 27-square-mile service area, part of Waymo’s broader plan to connect Silicon Valley and San Francisco

@Waymo: We’ve made millions of miles of memories over the past 15+ years, but today is special. We’re returning to where the journey began, gradually opening our doors to our first public riders in Mountain View, Los Altos, Palo Alto, and parts of Sunnyvale.

⚡️

@Waymo: We’ll gradually add riders to this new service area over time, as we work to bring Waymo One to more of the Bay Area. Keep an eye on the app if you’re an area local. Who’s ready to ride?

This rollout uses fully electric Jaguar I-Pace vehicles equipped with Waymo’s fifth-generation self-driving technology. 

+ Related content:

• @CNBCtelevision: Google's valuation ignores Waymo's self-driving success.

☑️ #177 Mar 7, 2025 🟠 opinion

A cost-effective alternative to cloud services like Google Cloud Platform

@rbranson: Just saved a customer $500,000 a year by moving them from GCP to this 4U server in my garage lol.

⚡️

https://x.com/rbranson/status/1898096536963498259 : update: due to overwhelming demand lol we are installing this new high end switch in the garage’s core infrastructure cabinet.

⚡️

@/rbranson: iykyk

☑️ #176 Feb 27, 2025

What if Doctors designed logos?

☑️ #175 Feb 19, 2025

The World's Largest Search Engine Doesn't Want You to Search

The Honest Broker: Why is Google abandoning its core mission?

The Honest Broker

The World's Largest Search Engine Doesn't Want You to Search

Almost everything in the digital world is turning into its opposite…

Read more

2 years ago · 2264 likes · 318 comments · Ted Gioia

The Honest Broker

A trustworthy guide to music, books, arts, media & culture by Ted Gioia

By Ted Gioia

​

💲 EARNINGS ANNOUNCEMENT: Feb 4,2025

Alphabet Announces Fourth Quarter 2024 and Fiscal Year Results (pdf)

Webcast

⇢ Some reactions(Before and After):

• @StockMKTNewz: Google $GOOGL has 5 separate businesses that make more than $30 Billion a year.

  

• @finchat_io: Google Cloud is a thing of beauty. Just gets more impressive with each passing quarter. $GOOG $GOOGL

​

☑️ #174 Feb 2, 2025

DeepSeekやo3-miniが話題ですが、Googleの発表した「DataGemma」も相当すごい

@tetumemo: [Translated from JP by Google] DeepSeek and o3-mini are hot topics, but Google's "DataGemma" is also pretty amazing.

In short, "The combination of RAG and RIG allows for accurate research with minimal hallucination (misinterpretation of facts)." What is RIG? What happens when you combine it with RAG? For details, please see the reply section ↓

+ Related content:

• blog.google (9/12/24): [Excerpt] DataGemma: Using real-world data to address AI hallucinations.

  DataGemma are the world’s first open models designed to help address the challenges of hallucination by grounding LLMs in the vast, real-world statistical data of Google's Data Commons.

  

  How Data Commons can help tackle hallucination

  As generative AI adoption is increasing, we’re aiming to ground those experiences by integrating Data Commons within Gemma, our family of lightweight, state-of-the art open models built from the same research and technology used to create the Gemini models. These DataGemma models are available to researchers and developers starting now.

  DataGemma will expand the capabilities of Gemma models by harnessing the knowledge of Data Commons to enhance LLM factuality and reasoning using two distinct approaches:

  • RIG (Retrieval-Interleaved Generation) enhances the capabilities of our language model, Gemma 2, by proactively querying trusted sources and fact-checking against information in Data Commons. When DataGemma is prompted to generate a response, the model is programmed to identify instances of statistical data and retrieve the answer from Data Commons. While the RIG methodology is not new, its specific application within the DataGemma framework is unique.

    • Example query: ''Has the use of renewables increased in the world?'' applying DataGemma RIG methodology leverages Data Commons (DC) for authoritative data.

  • RAG (Retrieval-Augmented Generation) enables language models to incorporate relevant information beyond their training data, absorb more context, and enable more comprehensive and informative outputs. With DataGemma, this was made possible by leveraging Gemini 1.5 Pro’s long context window. DataGemma retrieves relevant contextual information from Data Commons before the model initiates response generation, thereby minimizing the risk of hallucinations and enhancing the accuracy of responses.

    • Example query: ''Has the use of renewables increased in the world?'' applying DataGemma RAG methodology showcases greater reasoning and inclusion of footnotes.

    ​

• datacommons.org: Data Commons brings together the world's public data, making it simple to explore > docs.datacommons.org > Economy:

  • Australian Bureau of Statistics

    • Australia Statistics

  • Brazilian Institute of Geography and Statistics (IBGE)

    • Brazil Bank of Statistical Tables

  • Central Bureau of Statistics (CBS), Israel

    • Israel Census

  • Central Statistics Office(CSO), Ireland

    • Ireland Census

  • Data One Org

    • One.org Debt

  • European Union (EU) Eurostat

    • Regional Statistics by NUTS Classification

  • Food and Agriculture Organization of the United Nations

    • FAO Currency and Exchange Rate

  • General Statistics Office (GSO), Vietnam

    • Vietnam Census

  • India Ministry of Statistics and Programme Implementation

    • Periodic Labour Force Survey

  • MarineRegions.org

    • Global Oceans and Seas

  • Mexico National Institute of Statistics and Geography(INEGI)

    • Mexico Census

  • National Institution for Transforming India.

    • India Poverty Status

  • OpenFIGI

    • FIGI

  • Portal Site of Official Statistics of Japan (e-Stat)

    • Japan Census

  • Reserve Bank of India

    • Database on Indian Economy

  • Statistics Canada

    • Canada Statistics

  • U.S. Bureau of Economic Analysis (BEA)

    • GDP by County, Metro, and Other Areas

    • Gross Domestic Product

    • Regional Price Parities by State and Metro Area

  • U.S. Bureau of Labor Statistics (BLS)

    • Consumer Price Index

    • Job Openings and Labor Turnover Survey

    • Labor Force Statistics from the Current Population Survey (CPS)

    • Local Area Unemployment Statistics (LAUS)

    • Quarterly Census of Employment and Wages (QCEW)

  • U.S. Census Bureau

    • American Community Survey (ACS)

    • County Business Patterns

    • Economic Census

  • U.S. Commerce Data Hub

    • Economic Development Administration (EDA)

  • U.S. Department of Housing and Urban Development (HUD)

    • Income Limits

  • U.S. Department of Labor (DOL)

    • USA Minimum Wages

    • Weekly Claims and Extended Benefits Trigger Data

  • U.S. Federal Reserve

    • Treasury Nominal and Inflation-Indexed Constant Maturity Series

    • US Federal Reserve H.15 Interest Rates

  • World Bank

    • World Bank Datasets

    • World Development Indicators

☑️ #173 Dec 19, 2024

ModernBERT

huggingface.co: [Excerpt] Finally, a Replacement for BERT.

Introduction

BERT was released in 2018 (millennia ago in AI-years!) and yet it’s still widely used today: in fact, it’s currently the second most downloaded model on the HuggingFace hub, with more than 68 million monthly downloads, only second to another encoder model fine-tuned for retrieval. That’s because its encoder-only architecture makes it ideal for the kinds of real-world problems that come up every day, like retrieval (such as for RAG), classification (such as content moderation), and entity extraction (such as for privacy and regulatory compliance).

Finally, 6 years later, we have a replacement! Today, we at Answer.AI and LightOn (and friends!) are releasing ModernBERT. ModernBERT is a new model series that is a Pareto improvement over BERT and its younger siblings across both speed and accuracy. This model takes dozens of advances from recent years of work on large language models (LLMs), and applies them to a BERT-style model, including updates to the architecture and the training process.

+ Continue reading

+ Related content:

• research.google and arxiv.org (10/11/18): [Abstract] BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.

  We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.
  BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).

☑️ #172 Dec 17, 2024

Infosys strengthens strategic collaboration with Google Cloud to boost enterprise AI innovation, establishes Center of Excellence

infosys.com: [Excerpt] Leveraging Infosys Topaz and Google Cloud technology, the center of excellence will foster co-innovation to deliver transformative AI-powered solutions.

Infosys and Google Cloud share a strong legacy of delivering innovative data analytics and AI solutions, driving customer success. Infosys has already enabled more than 60,000 employees on Google Cloud. Combining the expertise of Infosys Topaz AI capabilities and Infosys Cobalt cloud capabilities, with Google Cloud's gen AI technology, the center of excellence will help address complex challenges across both business and technology domains using innovative enterprise AI solutions. The collaboration will enable a conducive environment for enterprises to co-create customized solutions in areas such as contact center AI, software development lifecycle, agentic AI, speech-to-speech, text-to-image, and application modernization, aimed at optimizing efficiency and business performance.

+ Continue reading

☑️ #171 Dec 17, 2024

Gemini Flash!

@rjvir: Google Gemini's market share among developers went from ~5% in September to >50% market share last week (per @OpenRouterAI)

+ Related content:

• openrouter.ai/rankings

  • gemini-flash-1.5-8b > 44.8M (12/9/24)

  • gemini-flash-1.5 > 35.8M (12/9/24)

• ai.google.dev: Gemini models

  • developers.googleblog.com (12/11/24): [Excerpt] The next chapter of the Gemini era for developers. Build with Gemini 2.0 Flash.

    Building on the success of Gemini 1.5 Flash, Flash 2.0 is twice as fast as 1.5 Pro while achieving stronger performance, includes new multimodal outputs, and comes with native tool use. We’re also introducing a Multimodal Live API for building dynamic applications with real-time audio and video streaming.

    Starting today, developers can test and explore Gemini 2.0 Flash via the Gemini API in Google AI Studio and Vertex AI during its experimental phase, with general availability coming early next year.

☑️ #170 Dec 13, 2024

Unlock the Infinite Possibilities of XR With Galaxy AI

news.samsung.com: [Excerpt] Available for both headsets and glasses, this new platform leverages the power of Gemini by bringing together a conversational user interface (UI) and a contextual understanding of the world around you. It offers cutting-edge capabilities beyond gestures or a controller, utilizing voice and natural conversation to elevate the experience. With these accurate, personalized interactions, the platform will be your helpful AI assistant.

Code named “Project Moohan”, the first headset designed for Android XR is poised to bring this experience to life in the near future. The name “Moohan”, meaning ‘infinity’ in Korean, connotes our belief in delivering unparalleled, immersive experiences within an infinite space. Equipped with state-of-the-art displays, Passthrough capabilities and natural multi-modal input, this headset will be your spatial canvas to explore the world through Google Maps, enjoy a sports match on YouTube or plan trips with the help of Gemini. All these experiences come with lightweight, ergonomically optimized hardware designed to ensure maximum comfort during use.

+ Continue reading

☑️ #169 Dec 10, 2024

Intersect Power forms strategic partnership with Google and TPG Rise Climate to co-locate data center load and clean power generation

intersectpower.com: [Excerpt] Collaboration accelerates the evolution of data center development and power provision, reducing costs, increasing reliability, and supporting local economic development. New ~$800M funding round led by TPG and Google will accelerate partnership’s development.

Under the terms of the partnership, Intersect Power will build new clean energy assets, with Google providing offtake via newly constructed data center campuses as an anchor tenant in co-located industrial parks. Once built, the Google data center would come online alongside its own clean power, bringing new generation capacity to the grid to meet its own load.

To further its vision throughout the U.S., Intersect Power also announced a more than $800 million funding round led by TPG Rise Climate and Google, with participation from Climate Adaptive Infrastructure and Greenbelt Capital Partners. Morgan Stanley & Co. LLC acted as financial advisor on the transaction.

+ Continue reading

+ Related content:

• therisefund.com: [Excerpt] About TPG Rise Climate.

  TPG Rise Climate is the dedicated climate investing strategy of TPG’s $25 billion global impact investing platform. TPG Rise Climate pursues climate-related investments that benefit from the diverse skills of TPG’s investing professionals, the strategic relationships developed across TPG’s existing portfolio of climate-focused companies, and a global network of executives and advisors. The fund takes a broad-based sector approach to investment types, from growth equity to buyouts, and focuses on climate solutions in the following thematic areas: clean electrons, clean molecules and materials, and negative emissions. Jim Coulter, TPG Founding Partner and Executive Chairman, serves as a Managing Partner of TPG Rise Climate. Former U.S. Treasury Secretary Hank Paulson serves as TPG Rise Climate’s Executive Chairman. For more information, please visit therisefund.com/tpgriseclimate

☑️ #168 Dec 9, 2024

AI computing capacity fore leading tech companies

@thexcapitalist: $GOOG is way ahead in AI race than you think.

It has the largest AI computing capacity among cloud providers.

What’s even better is that 70% of its capacity comes from its own TPU chips.

$MSFT and $AMZN are completely dependent on Nvidia.

I can’t imagine being bearish

+ Related content:

• epoch.ai/research

☑️ #167 Dec 9, 2024

Google Willow

@sundarpichai: Introducing Willow, our new state-of-the-art quantum computing chip with a breakthrough that can reduce errors exponentially as we scale up using more qubits, cracking a 30-year challenge in the field. In benchmark tests, Willow solved a standard computation in <5 mins that would take a leading supercomputer over 10^25 years, far beyond the age of the universe(!).

⚡️

@sundarpichai: We see Willow as an important step in our journey to build a useful quantum computer with practical applications in areas like drug discovery, fusion energy, battery design + more. Details here:

https://blog.google/technology/research/google-willow-quantum-chip/

+ Related content:

• @GoogleQuantumAI: Meet Willow, our state-of-the-art quantum chip.

• @datascienceinyourpocket: Google Willow : Quantum Computing has changed forever.

☑️ #166 Dec 5, 2024

Waymo: 175,000 rides last week

@joecarlsonshow: So many people believe Waymo can't scale. Here are the numbers:

• 50,000 rides a week 6 months ago.

• 100,000 rides a week 3 months ago.

• 150,000 rides a week 1 months ago

• 175,000 rides last week.