NVIDIA’s $3.5 billion subscription to MediaTek’s convertible bonds can be read as a vote of confidence in one of Asia’s strongest fabless chip designers. NVIDIA is using its balance sheet to shape the market for custom XPUs before that market can develop around rival or open scale-up fabrics.

Hyperscalers increasingly want workload-specific accelerators that can improve performance per watt and total cost of ownership. Those chips can displace some general-purpose GPU demand. NVLink Fusion is NVIDIA’s answer. It allows customers to build differentiated accelerators while retaining NVIDIA’s interconnect, memory, networking, rack design and software compatibility. MediaTek supplies the custom silicon, high-speed I/O, packaging and manufacturing integration; NVIDIA keeps the system-level control point.

The financing terms are unusually attractive to MediaTek: a five-year, zero-coupon instrument issued at par, with conversion set 15% above the August 31 close and maximum dilution of about 1.67%. NVIDIA subscribed for $3.5 billion of the $3.9 billion issue. It receives no voting interest unless it converts and no cash interest while it waits. The initial share price only needs to compound about 2.8% annually to reach the conversion strike, but NVIDIA’s true return hurdle is higher because yielding cash has an opportunity cost.

1. What NVIDIA Is Actually Buying

Read more

The strongest version of the thesis is not that Bitcoin will eliminate gold, replace the dollar or become the sole reserve asset. It is that a scarce, neutral, internet-native bearer asset can earn a meaningful share of global demand for wealth that is held outside any single government’s liability structure.

2. The right valuation framework is gold—not a discounted cash-flow model

Read more

Read more

Nvidia will lend its balance sheet — indirectly, through guarantees and residual-value support — to companies that cannot otherwise borrow at rates which make buying Nvidia hardware economic.

What exists today is non-binding. No committed capital, no identified borrower, no pricing, no collateral package. Nvidia has said it may provide residual-value support covering up to 25% of a given opportunity, without defining the trigger or the loss layer. Its April 2026 Form 10-Q discloses $3.5 billion of facility lease guarantees. That is the only fully quantified guarantee on the books.

Vendor financing is nothing new and can cut two ways - it can be very profitable and strategic or it can backfire and stretch a company to the brink. In this note we explore what this strategic move means and what we can learn from history on the soundness of it all.

Read more

1. The scale of what has been committed

Amazon, Alphabet, Microsoft and Meta have guided to roughly $630–725 billion of combined capital expenditure in 2026, up from about $390–410 billion in 2025. Add Oracle, the neoclouds and new entrants and 2026 AI-attributable capex lands near $850 billion. Consensus for 2027 sits above a trillion.

The least-discussed entrant is SpaceX. Following the xAI combination it disclosed $18.4 billion of capex in the second quarter of 2026, of which $15.8 billion was AI — roughly six times its connectivity and space segments combined, and more than double the prior quarter. It is funded by a Starlink business reporting $7.81 billion of quarterly revenue, up 92% and ahead of consensus; the shares fell anyway, on the capex line. SpaceX has also filed with the FCC for up to one million orbital data-centre satellites, with compute payloads targeted from 2028.

Chart 1 — AI-attributable capex by player. Source: company guidance and disclosure; 2027E author estimate.

Two features matter more than the level. Spending is accelerating, not merely high, so the hurdle is a moving target. And the composition has changed: in 2023 this was four cash-rich incumbents self-funding; by 2026 roughly a quarter of the spend comes from entities that must raise external capital. Global AI-related debt issuance is on track for near $570 billion this year, with an estimated $800 billion of data-centre financing in private credit and off-balance-sheet vehicles.

2. What an acceptable return requires

Read more

Three pieces of evidence arrived together.

Read more

It is also right often enough to be useful. The honest position is that wrapper is not a category — it is a spectrum, and the work is figuring out where a company sits on it. What follows: the architecture of an LLM-powered application, what changes when you depend on one model versus several, what is and is not copyable, ten AI-native companies scored against that framework, and the two escape hatches the best of them are already using. I do not think the app layer is doomed. I also do not think most of it is defensible. Both can be true, and the whole job is telling them apart.

Read more

Despite some strong design wins with Samsung and other manufacturers, I still see some challenges with price/performance positioning especially for PCs which the company recognizes and is addressing. But the interesting thesis point for Qualcomm and the sector is how much compute will move to local AI, low-power compute and persistent connectivity that will allow Qualcomm and others to sell more silicon content to the same user.

Investor Day labeled this strategy $6 billion of fiscal 2029 “Personal AI and Compute” revenue inside a greater-than-$14-billion IoT target. Yet the number is mostly a repackaging: Qualcomm’s 2024 goals were $4 billion for PCs and more than $2 billion for XR. The combined bucket also reduces disclosure granularity. PCs and XR sit inside IoT while Samsung phone content sits in Handsets; those QCT categories generated $1.73 billion and $6.02 billion in fiscal Q2 2026, but the businesses examined here are not disclosed separately.

Figure 1. Qualcomm’s opportunity is cross-device attach, not a single product category.

PCs: from proof of concept to a real product ladder

Read more

Read more

See our 30 page industry report for Paid subscribers to have a different look at industry dynamics, historical analogies, where we are in different parts of the cycle and who the winners and losers may be.

Read more

IBM’s preliminary second-quarter results triggered a 25% stock decline because the warning cut directly across the market’s core debate: is generative AI expanding enterprise technology budgets, or merely consuming funds that would otherwise have gone to traditional software and services? The immediate answer is both. IBM reported $17.2 billion of revenue versus $17.86 billion expected, a 3.7% miss, while adjusted EPS of $2.93 was 2.7% below consensus. Those headline gaps were manageable. The more important signal was the composition: software grew only 5% versus roughly 10% expected, and infrastructure declined 7% versus an expected decline of approximately 3%.

The best interpretation is a two-part shock. First, enterprise customers are reallocating near-term budgets toward servers, storage and memory as AI demand creates supply constraints and price inflation. Second, IBM was unusually exposed to that shift because the company expected a stronger z17 mainframe cycle and associated transaction-processing software, then failed to close multiple large deals late in June. Management said those missed deals drove the majority of the shortfall. Red Hat still grew 11%, which is inconsistent with a broad collapse in hybrid-cloud and modernization demand.

INVESTMENT CONCLUSION

Read more

1. The rate-of-change, in one chart

Read more

The proprietary US frontier lead has compressed from generations to single-digit benchmark points. DeepSeek, Zhipu (GLM), Moonshot (Kimi), and Alibaba (Qwen) now field open-weight models within roughly six to fourteen composite points of the best closed US systems, and lead on cost by one to two orders of magnitude. Europe has not closed the frontier gap in absolute capability terms, but Mistral has built a viable second tier and a sovereignty-driven distribution channel that does not require frontier parity to monetize. The practical effect is a bifurcating market:

US labs are retreating upmarket into agents, long-context reasoning, and enterprise workflow lock-in, while Chinese open weights commoditize the base layer globally, particularly in cost-sensitive and non-aligned geographies. Capital allocation follows distribution and balance-sheet depth more than raw model quality. The named winners below

Read more

Three device-layer transitions are converging in 2026: a rumored OpenAI always-on AI hardware product that would sit in the background and respond through voice, cameras, and context — now confirmed delayed to 2027 — a disputed but strategically coherent SpaceX/xAI handset concept, and an accelerating land-grab in AI smart glasses led by Meta. Layered underneath is a satellite broadband capacity race between Starlink and Amazon Leo that is no longer just a connectivity story — it is becoming AI distribution infrastructure. SpaceX’s February 2026 acquisition of xAI converts Starlink from a bandwidth business into a candidate full-stack AI platform: model (Grok), pipe (Starlink Mobile/Direct-to-Cell), and prospective device. No other AI lab has all three layers under one roof.

Our framework below separates rumor from confirmed fact at each step, then maps winners and losers across silicon, hyperscalers, eyewear/OEM partners, and orbital infrastructure.

This is a mapping exercise ahead of Q3 catalysts

Read more

Technology finished the first half as one of the market’s clear leadership groups, but the useful observation is narrower: the winners were not the same companies investors owned for the 2023-25 AI trade. Public-market scoreboards showed the S&P 500 up roughly 9.5% and the Nasdaq up low-teens through late June, while semiconductor proxies were up multiples of that. The PHLX Semiconductor Index was up about 88% for the half, and the top-performing S&P 500 list was dominated by technology hardware, memory and AI infrastructure names.

In this note we present a short 3 min video overview followed by a more detailed industry and sector piece. We look at both 1H26 performance as well as scenarios and our views for the rest of 2026 into 2027.

Read more

Read more

Here were the five things I thought were most interesting: the data-center push and where the advantage really is; whether Qualcomm is positioned for inference moving to the edge; whether “distributed inference” is a real trend or a convenient narrative; what 6G actually changes for agentic AI; and why management felt confident enough to take long-term guidance up.

Read more

“Unmetered intelligence” means AI inference becomes local, abundant, and cheap enough that users stop thinking of every prompt as a billable cloud event. Instead of renting intelligence token-by-token from a remote data center, the PC becomes an always-on reasoning layer close to the user’s files, apps, workflows, and private context.

Microsoft’s framing around Windows is important: build apps with local AI for “unmetered intelligence” on every PC. The ambition is to shift AI from a scarce cloud service into a native operating-system capability — embedded in the device, callable by applications, and available when cloud inference is too expensive, slow, or privacy-sensitive.

What “Unmetered Intelligence” Means

AI has mostly been consumed as a metered cloud product: every prompt, code completion, image request, agent step, or enterprise workflow carries token, compute, network, latency, and margin costs. Even when users do not see the bill, someone pays for inference.

Unmetered intelligence flips the model. It does not mean AI is free; it means the marginal cost of many everyday tasks falls low enough that intelligence feels like a built-in resource. Like broadband replacing dial-up, local AI could make reasoning feel ambient rather than rationed.

Practically, this means small and specialized models run locally; larger models are invoked only when needed; workloads route dynamically across CPU, GPU, NPU, and cloud; and applications are designed around continuous assistance rather than occasional chatbot use.

Why the PC Matters Again

The PC matters again because the most valuable work context already lives on or near the device: documents, browser sessions, emails, spreadsheets, meetings, code repositories, calendars, screenshots, and application state.

Sending all of that context to a cloud model is expensive, slow, and sensitive. Windows can become more than an app launcher: an agentic runtime that hosts models, exposes secure APIs, manages permissions, and lets developers use local inference by default.

The cloud will still train and run frontier models, but the endpoint can handle high-frequency tasks. The data center remains essential; the PC becomes where AI is constant, personalized, private, and cheap enough to use all day.

How It Works

1. Local models: Small language models, vision models, speech models, OCR, embedding models, and task-specific agents run directly on the PC. These handle fast, repetitive, privacy-sensitive tasks.

2. Heterogeneous compute: The operating system routes workloads to the right processor: CPU for general tasks, GPU for parallel heavy inference, NPU for efficient background AI, and cloud for frontier-scale reasoning.

3. Developer APIs: Instead of every software vendor building its own AI stack, Windows exposes APIs for local inference, model optimization, and application integration.

4. Agent permissions and security: If AI agents can act across apps, the OS must enforce identity, containment, consent, logging, and enterprise controls.

This is the difference between a chatbot and a platform: a chatbot answers questions; a platform lets every application become AI-native.

What It Enables

· Personal knowledge retrieval: Search across files, screenshots, browser history, and meetings using natural language without pushing every query to the cloud.

· Writing and editing: Draft, summarize, rewrite, translate, and format documents locally, with cloud escalation for higher-quality reasoning when needed.

· Developer workflows: Code completion, debugging, test generation, documentation, local agents, and terminal assistance that run close to the project context.

· Creative production: Image enhancement, video editing, background removal, generative effects, and media indexing on-device.

· Enterprise automation: Agents that operate within approved application boundaries, summarize internal knowledge, and complete repetitive workflow steps with lower data-exfiltration risk.

The key is frequency. A cloud chatbot may be used dozens of times a day; local intelligence can be used hundreds or thousands of times because cost, latency, and privacy friction are lower.

The Economics: A Capex-for-Opex Swap

Read more

Microsoft Build 2026 was not primarily a feature-release event. It was a platform-control event. Microsoft used two days at Fort Mason Center in San Francisco to show how it intends to own the enterprise AI agent stack from the developer workflow through the data layer, runtime environment, governance plane, model layer, device endpoint, and cloud execution infrastructure. Satya Nadella’s opening frame — “AI alone won’t change your business. The system running it will” — was not a tagline. It was a product roadmap.

The core message is that agents need more than models. They require trusted context, governed access to internal and external data, secure execution environments, identity, observability, deployment surfaces, and cost-efficient compute. Microsoft’s answer is an integrated stack: GitHub Copilot for developer workflow, Microsoft IQ for organizational context, Foundry for agent development and production operations, Fabric and databases for data grounding, Windows and Windows 365 for local and cloud agent execution, Entra/Intune/Defender/Purview for governance, and Azure infrastructure for inference at scale.

The strategic read-through is positive because MSFT is positioning as the default enterprise control plane for agentic AI, not merely a reseller of frontier models. The announcement set also demonstrates Microsoft reducing dependency risk by expanding in-house models (seven new MAI models, zero OpenAI distillation), open-source runtimes, NVIDIA partnerships, and Fabric/database integration. The financial read-through is less immediate — Build clarified product direction but did not by itself prove incremental revenue, Copilot attach acceleration, or AI gross-margin stability. Treat Build as strategically validating but not yet estimate-changing.

As with all major developer events like this, Microsoft not only revealed its future roadmap, but also gave directly and indirectly gave us a window into potential longer term winners alongside Microsoft as well as potential losers which we will highlight in this note.

Read more

The consensus story is artificial intelligence wants more silicon than the world can currently build. CoWoS packaging is sold out. High-bandwidth memory is sold out. Leading-edge foundry slots are allocated into 2027. Every quarter a CEO says the same magic words — demand continues to far exceed our supply — and the stock goes up.

This report asks a more uncomfortable question. If a sector physically cannot satisfy demand, what actually happens to its earnings estimates over the next few quarters? And what happens to a stock priced for those estimates to keep climbing — particularly the momentum money that bought the revision trend, not the business — when the revisions flatten against a wall everyone can already see?

The answer is a paradox the most crowded names in the SOX may be walking into: the harder a sector is constrained, the less room it may have left for the upside earnings surprises that justify its multiple. You cannot beat estimates on volume you cannot make.

1. A sentence that means the opposite of what it sounds like

Start with the arithmetic, because everything downstream falls out of it. A chipmaker’s revenue is, to first order, units shipped × average selling price. When a company is genuinely capacity-constrained, the “units” term is not a function of demand at all — it is a function of how many wafers, stacks, or packages the factory can physically push out the door. Demand can triple; if the line runs at 100% utilization, shipments don’t.

That leaves only two levers that can push reported revenue — and therefore consensus estimates — higher than what the Street already models:

• Scheduled capacity additions. New fab shells, bonding tools, packaging facilities. Real — but disclosed in advance, 12–24 months to qualify, and already in everyone’s model. Knowable, not surprising.

• Price. The only fast lever. But price has customer pushback, is the part of the equation that historically mean-reverts, and is the first thing a downturn takes back.

Momentum is a derivative. It doesn’t feed on how big earnings are — it feeds on how fast estimates are still rising. A hard capacity ceiling attacks exactly that second derivative.

Here is the trap in one move. When a constrained company reports a blow-out quarter and says “demand far exceeds supply,” but does not raise its disclosed capacity number, there is nothing for the sell-side to revise upward. The ceiling is already in the model. The print is great, the narrative is euphoric — and the estimate revision is zero. For a stock whose multiple was paying for continued upward revisions, zero is a disappointment. That is the mechanics of “sell the news” in a supply-constrained group.

2. Mapping the constraint: where the AI chain is actually tight

The single most important fact of the 2025–2026 cycle is that the bottleneck moved. The 2023 shortage was about wafers. By 2026 the wafer is the easy part; the constraint migrated downstream into advanced packaging and memory — a handful of specialized back-end processes the whole industry now funnels through.^[1][5] A wafer without CoWoS is not an AI accelerator; it is silicon waiting in a queue.

Read more