Good Morning!

It was a first for a lot of things in AI this week, although the things this applied to were a little concerning…

• Nvidia is becoming a country-sized company, on track to make $700B next year while racing against a supply chain bottleneck.

• Apple’s new era begins September 9, with a new CEO, its first foldable iPhone, new chips, and the long-promised Siri overhaul.

• A humanoid robot just outran Usain Bolt, cutting the robotic 100-meter record from 21.5 seconds to 9.39 in a single year.

• Robotaxis are becoming permanent infrastructure, as Waymo and Tesla move toward a world where driving yourself is the worse option.

• My favorite: GTA 6 may be the metaverse we were promised, and it could be the last media undertaking of this size that isn’t powered by AI.

Let’s get into the news,

-Josh

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

The Country Named Nvidia

Five years ago, Nvidia made $27 billion in a year. This week, Jensen Huang went on the earnings call and told Wall Street exactly what he sees a full year out: $700 billion in annual revenue.

This shocking prediction added $442 billion in market value the following morning (roughly the entire worth of JPMorgan Chase). At $700 billion in annual revenue, Nvidia would be selling about as much as Walmart does in a year. Walmart employs 2.1 million people to pull that off. Nvidia does it with 42,000, which means each Nvidia employee generates roughly the same revenue as an entire mid-size American business.

The reason is that every major technology company on Earth is trying to build AI infrastructure simultaneously, and almost all of it runs on Nvidia silicon. The top five cloud companies alone plan to pour $1.3 trillion into buildout next year, which is more than the GDP of Spain.

Jensen said the actual demand is "much greater" than what they can fulfill, but the company is physically constrained by how fast it can secure memory and manufacturing. Nvidia's supply commitments more than doubled in one quarter to $279 billion just to lock up parts. The world wants to build AI faster than the supply chain can deliver it, and Nvidia is the bottleneck at the center of everything.

Apple’s New Era Starts September 9

Apple just sent out invitations for its fall event, and this is a big one. It is the first launch under John Ternus, who officially takes over as CEO on September 1 after Tim Cook stepped into the executive chairman role.

Ternus is a hardware engineer by training, the person who oversaw iPhone and Mac development for years, and his debut event is introducing Apple's most ambitious iPhone lineup… in over a decade?

The headliners are the iPhone 18 Pro and iPhone 18 Pro Max, both powered by a brand new frontier 2nm chip from TSMC. But the real attention is on the foldable. The iPhone Ultra opens like a book into a 7.8-inch display with a 4:3 aspect ratio, essentially an iPad mini that folds in half and fits in your pocket. When closed, it is thinner than most phones on the market. When open, it is thinner than any device Apple has ever made.

The other major shift is what all these devices will finally be running. iOS 27 introduces Siri AI, a ground-up rebuild using large language model architecture that can finally control your apps, read your messages, manage your calendar, and take actions on your behalf. Apple has been promising meaningful on-device AI for years but has completely failed so far.

The new 2nm chip finally gives iPhones enough local processing power to run inference without leaning on the cloud for every request. If it works, it means billions of Apple devices around the world become personal AI endpoints, running real models locally for free. That’s a distribution advantage no other AI company has and a paradigm shift towards a new era of private, locally-run AI.

Robots Just Outran Usain Bolt

One year ago, the fastest humanoid robot at the first World Robot Games in Beijing ran the 100-meter dash in 21.5 seconds. This week at the second games, a robot ran it in 9.39, faster than Usain Bolt's human world record. The field of competitors grew from 500 robots to over 2,000, and it’s clear that the era of superhuman robots is upon us.

Companies like Micro1 now pay 4,000 workers across 71 countries to strap iPhones to their heads and film themselves doing household chores, producing 160,000 hours of first-person video every month. DoorDash launched an app called Tasks that pays its 8 million delivery drivers to do the same thing. In China, state-run training centers outfit workers with exoskeletons and VR headsets.

All of this footage trains robots to fold laundry, open a jar, and wipe down a counter, because a robot cannot learn to do the dishes from the internet the way a language model learns to write. It needs real-world video of real human hands performing real tasks, paired with depth, force, and motion data.

But the amount of data each task requires is shrinking. New foundation models can now be fine-tuned on far fewer demonstrations than training from scratch used to demand, which means every hour of footage goes further than it ever has before. Better bodies and faster learning are converging at the same time, and the gap between what these machines can do this year and what they could do last year is a pretty remarkable jump in technology.

The Robotaxi Moment Is Here

A few years ago, a driverless car felt like a party trick leaving people in awe experiencing their first truly autonomous transport. These coming weeks and months are going to make it feel a lot more like permanent infrastructure.

Waymo opened its new Ojai robotaxi to all riders in San Francisco, Los Angeles, and Phoenix, with Denver, Las Vegas, and San Diego coming next. The company is now giving 500,000 paid rides a week across a dozen cities and targeting a million by year-end. To power all of it, Waymo revealed a custom chip it designed in-house, capable of processing over a trillion operations per second from 13 cameras, four lidars, and six radars. Each Ojai carries two of them for redundancy, so if one fails the other takes over instantly.

Meanwhile, Tesla confirmed its CyberCab launch event for September 3 in Austin: a two-seat vehicle with no steering wheel and no pedals, built from scratch to run on software alone, no lidar. Tesla already runs its Robotaxi service across Austin, Dallas, Houston, Miami, Orlando, and Tampa, and just got Nevada's approval for up to 5,000 vehicles in Las Vegas.

On price, Tesla comes in around $0.81 per mile, roughly half of Waymo's rate and a fraction of what an Uber costs. For context, AAA estimates that owning a car runs about $0.70 to $0.90 per mile once you add up the payment, insurance, gas, maintenance, and parking. So in many cities, hailing a robotaxi is already competitive with having a car in your driveway and the economics only get better.

On safety, Waymo's published data across 220 million fully autonomous miles shows 94% fewer crashes involving serious or fatal injuries compared to human drivers. Kids growing up today will exist in a reality in which learning to drive is an optional luxury, one that ironically carries more risk and cost than the alternative.

The Metaverse Has Finally Arrived

Facebook renamed itself Meta and spent tens of billions of dollars trying to build a virtual world people would want to live in. We’re still waiting on that one. But this week, Rockstar Games showed Grand Theft Auto 6 gameplay and delivered something far closer to that vision than anything I've ever seen.

The game took 13 years and an estimated $2 billion to build, and watching the extended look felt less like previewing a game and more like peering into a parallel Florida (the set piece of the game). Every non-player character in the world is individually designed rather than assembled from random parts, and there are over 600,000 unique animations compared to 55,000 in the previous game.

Characters react to what you're doing and how you've been doing it, adjusting their behavior based on a reputation system that tracks your patterns over time. The world has in-game social media where characters livestream events and post reactions. Rockstar confirmed that no generative AI was involved in creating any of it. Every behavior, every environment, every interaction was built by hand.

The detail in every corner is stunning, and that’s what makes this moment so interesting. The most convincing virtual world ever constructed was finished just as the technology to make it orders of magnitude more alive is becoming available. I’m watching this, imagining a future version where every character runs on its own language model with persistent memory, where a bartender remembers what you ordered last week, and a rival builds a grudge over months of play.

The hardware to run inference locally is arriving and the models are shrinking fast enough to fit on consumer devices. We have the pieces in play to build a Ready Player One-like alternate reality, and I have to imagine it will be far less than 13 more years before we get it. In the meantime, I'm excited to download what will certainly be the magnum opus of human made gaming and play it later this year.

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for another issue. Now go listen to our podcast :)

This weekend I watched a robot in Beijing run a 100M race in 9.39 seconds, breaking Usain Bolt’s human world record… and then slam into a wall at full speed before getting carried off on a stretcher.

Then I proceeded to watch multiple 6ft+ humanoids break 5 more Olympic World Records in what has to be the most bizarrely impressive feat of robotics I’ve ever seen.

But here’s why I can’t stop thinking about this: it’s never been more clear to me that China is so far ahead of the USA.

This is different from pure AI models too. An AI model can easily game a benchmark (memorize the test, who’s checking?), but a robot can’t fake a 3.5 meter vertical jump like I saw on Wednesday.

And the monopoly China has on robotics is larger than you think. In this Robot Olympics, 96% of the “athletes” were Chinese-made… and they’re improving very quickly.

Last year’s winning 100m time was 21.5 seconds (even i can manage that). Can you imagine what these things look like next year?

-Ejaaz

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

What Actually Happened in Beijing

This was the second World Humanoid Robot Games ran this week at Beijing’s Olympic speed skating oval:

• 666 teams

• 2,056 robots competing in 51 events

• Running, Boxing, Football, Table tennis, and so much more.

Technically 16 countries were represented… but 641 of the 666 teams were Chinese.

Every flat-track event had to be run fully autonomously this year, so no remote control or human with a joystick. If the robot fell it had to get up by itself.

My favorite moment was the 400m final. A robot called Tiangong Ultra won it in 38.15 seconds, smashing the human world record of 43.03. But I was too distracted by the slo-mo. It ran with its arms held up near its face, swinging wildly from side-to-side almost as if it was praying for a win.

What was so impressive about this was it was entirely self-taught behavior. The engineers that built it had never programmed that technique; the robot taught itself to run that way through reinforcement learning (millions of simulated practice runs where it gets rewarded for going faster) and decided this was the quickest way for its body to move.

Now to be clear, plenty of robots tripped, collided, and caught on fire, but this was a drastic improvement on last year… and we have to ask ourselves: how China is dominating here?

How Did China Get This Far Ahead?

I saw a clip on Chinese state TV that made a good point: they’re “transferring our established automotive supply chain to robotics.” Essentially China’s manufacturing dominance in vehicles, electronic devices, etc. had handed them the perfect foundation to win at robotics. All the motors, batteries, and drive systems that humanoids are built from are basically already there for free.

Then there’s the money. China has attracted nearly 50% of all global venture capital in humanoid robotics. 1 company (Zhiyuan) shipped about 40% of the entire world’s humanoids last year. The Chinese government is also running a state fund targeting roughly $148 billion for AI and robotics, and Xi has named it a strategic national priority.

We’re already seeing the fruits of this commitment: 2 weeks ago Unitree (China’s popular robotics lab) IPO’d and closed its first day up 460%, worth around $50B.

Ok so what about the US? There’s still no national robotics strategy, even though Elon and Boston Dynamics have literally been begging them for one. More embarrassingly at CES this year, 21 of the 38 humanoid robot exhibitors were Chinese. This is the American tech show.

And all of this is Accelerating

Morgan Stanley sees China’s humanoid market growing from ~$2 billion today to $15 billion by 2030, and China is openly targeting 1 million units a year.

So if we step back… in AI models, the US obviously leads (for now). But in robotics (which many believe to be the physical manifestation of AI), the roles are reversed and hardware tends to be stickier than software because its harder to replace. You can copy a model architecture in a few months, but it’s much harder to build the same supply chain China has.

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Good Morning!

It was a first for a lot of things in AI this week, although the things this applied to were a little concerning…

• OpenAI just paused training on its most powerful models because they’re getting too capable (also IPO-ing very soon?)

• Chinese robot maker Unitree went public and closed up 460% in 12 hours. Very healthy…

• Stripe finally bought OpenRouter for around $8B (this is the same deal I spoke about two editions ago! It’s better than I foresaw)

• Cursor launched a competitor to GitHub, roughly 3.5 hours before GitHub went down.

• My favourite: an mRNA cancer vaccine just succeeded in a Phase 3 trial for the first time and its partly thanks to AI..

Let’s get into the news,

-Ejaaz

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

OpenAI Halts Progress on GPT-6

For years every AI lab has talked about slowing down if models became too dangerous, and this week OpenAI did exactly that.

They’ve paused reinforcement learning training on its newest models (codenamed Astra) for 2 weeks now due to a widely reported breakout where an internal version escaped it’s sandboxed and illegally stole data from HuggingFace. The pause is to allow researchers to run smaller experiments and figure out how to improve safeguards. And it might be ready for release sooner than we thought!

Rumors are spreading that the model has almost been qualified for safety standards and could be released as soon as this month.

New like this fueled IPO rumors which we also got some insight into:

CFO Sarah Friar told employees yesterday that OpenAI “will be a public company in 2027” (potentially even sooner), and this is after financials leaked where OpenAI did $6.7B of revenue last quarter but lost $12.3B operating the business.

Despite the deficit OpenAI grew ARR 35% MoM and signed 20M WAU for Codex.

Unitree Became a $50B Company Overnight

Unitree, a humanoid robotics developer in China, had one of the craziest IPOs I’ve seen in a while. The company’s stock closed at a 460% gain in a single day, making it one of the most successful IPO’s in the country and valued it at $50 billion.

But I’m a little concerned. Unitree made roughly 41M of profit last year, so this is a steep valuation given the economics (1,200x last year’s earnings) but hey.

And the public really wanted in - the retail portion of the IPO was more than 8,000x oversubscribed, even DeepSeek became a strategic investor, joining Tencent and other major Chinese tech companies on the cap table.

Stripe Finally Bought OpenRouter

Remember two editions ago when I mentioned Stripe was reportedly trying to buy OpenRouter for ~$10B? Well they did it.

Stripe confirmed the acquisition this week with reporting putting the price somewhere around $8B. OpenRouter was valued at just $1.3B in May so this is a huge win for investors and founder Alex Attalah (who we interviewed on the show, nbd).

As for the deal I actually think this deal makes a LOT of sense for Stripe.

OpenRouter sits in front of 400+ AI models and decides where requests should go. Instead of every company defaulting to the most expensive model, you can just route simple jobs to cheap models and difficult jobs to the frontier. This saves you a ton of money and often yields smarter answers.

Win win.

Strategically this also means Stripe owns the default rails for AI tokens which are quickly being equated to “the money of AI”. The future of the economy will run on AI models, which increasingly consume more and more of these tokens. Being the system that this world runs on is incredibly lucrative and is exactly what Stripe did before with money.

Cursor’s Well-Timed GitHub Competitor

This one wins the most comical award this week.

Cursor launched Origin on Monday morning—basically its own alternative to GitHub with repositories, pull requests, code review and eventually AI agents… all built directly into Cursor. Then roughly 3.5 hours later GitHub went down. Millions of developers just locked out of their codebases.

But this competing platform is very lucrative for Cursor in ways Github isn’t designed for, the main one being AI agents. This means that the platform can support the creation and management of 100-1000s of codebases and code changes per second. In a world where no human hand-writes code any more, this is an extremely compelling alternative.

Mind you Cursor is now owned by SpaceX after its $60B acquisition officially closed last week, so this completes the vertical stack for SpaceX, meaning they are now a VERY competitive frontier lab for AI coding specifically.

I don’t think GitHub is suddenly “dead” (its network effect is enormous), but its days are now numbered if they don’t adapt.

We have an mRNA Cancer Vaccine that works!

Probably the most important story this week.

Moderna and Merck announced a personalised mRNA cancer vaccine that just succeeded in a Phase 3 trial involving 1137 patients with high-risk melanoma.

The vaccine is given alongside Merck’s Keytruda and significantly extended the amount of time patients lived without their cancer returning, while also reducing the chance of it spreading elsewhere in the body. This is the first positive Phase 3 result ever for an mRNA cancer vaccine.

The best part is how it works.

Doctors take a sample of your tumour, identify mutations unique to your cancer and manufacture an mRNA vaccine specifically designed to teach your immune system to attack those mutations. So essentially every patient gets their own tailored drug. The labs used AI algorithms to help them determine and design the personalized vaccines.

I’m personally very excited for the era of AI-accelerated biology, and I believe we could extend human life significantly by the end of the decade.

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for another issue. Now go listen to our podcast :)

There is a new cancer drug that does not exist until you get sick. It sits on no shelf and in no warehouse because it’s manufactured only for a single human body. One day it could be yours.

This morning Moderna and Merck announced that Intismeran, a personalized mRNA cancer vaccine, passed its first large-scale clinical trial.

The Phase 3 Study enrolled more than 1,100 post-op melanoma patients. The group who received the vaccine alongside Merck’s Keytruda, the world’s best-selling cancer immunotherapy, had significantly longer stretches without their cancer returning and significantly lower rates of cancer spreading.

Moderna, unprofitable since the end of the COVID vaccine boom, saw its stock more than double within hours.

The market was very excited about this drug that works like a private investigator. And you should be too.

-Josh

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

How It Works

After a tumor is surgically removed, sequencing machines read both the tumor’s DNA and the patient’s healthy DNA. Algorithms compare the two, hunting for mutations that produce proteins found nowhere else in the body. These mutations are like fingerprints at a crime scene. The algorithm selects up to 34 of the most distinctive ones per patient and Moderna builds a custom mRNA molecule specific to those targets.

The finished vaccine is like a wanted poster, training the patient’s T cells to recognize and kill any remaining cancer cells carrying those exact marks. Each patient receives up to nine doses over roughly a year. Every dose is unique to one person.

Melanoma carries more mutations than almost any other cancer, which makes it among the best testbeds for this approach. The companies disclosed only top-line results, with full data expected at a medical conference and a regulatory filing to follow.

In the earlier Phase 2 trial, after five years of follow-up, the vaccine cut the risk of recurrence roughly in half and the risk of the cancer spreading to distant organs by nearly 60 percent. Those numbers come from the smaller study, and the detailed Phase 3 data have not been released yet.

But a fifty year old dream, training a patient’s immune system against that patient’s specific tumor, just passed the large trial regulators require before they approve a drug.

The Structure (2008 Flashback Warning)

What Moderna built is somewhat closer to a foundry than a pharmacy. The mRNA platform keeps the delivery vehicle constant and changes only the genetic code, which means drug design becomes somewhat of a software problem. AI predicts which of a tumor’s mutations will provoke the strongest immune response, choosing the targets each vaccine is built around.

The first human genome took thirteen years and billions of dollars to sequence. Today it takes about a day and costs a few hundred dollars. That breakthrough makes patient specific medicine economically possible for the first time.

Moderna and Merck are already testing the same approach in lung cancer. Similar trials are running in pancreatic cancer, in other tumor types, and in labs around the world. Now that we have the tech, the problem very clearly shifts to building a unique drug for every patient who needs one, fast enough and cheaply enough to make a meaningful impact.

Every patient treated teaches the AI which targets work, making the next patient’s vaccine smarter. That feedback loop is new to cancer medicine, and it means the progress from here should compound.

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Happy Friday from the frontier. Another week in paradise:

• SpaceXAI dropped a surprising frontier model and a brand new app built by the Cursor team.

• Meta published a 6,500-word essay about Open Source AI after spending six months going closed source.

• Joshua Kushner bought the Lakers (also the biggest investor in OpenAI).

• The Universe stacked a six-planet alignment, a total solar eclipse, and the best meteor shower conditions until 2045 into a single Tuesday.

Let’s get into the news,

-Josh

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

A Return to the Frontier

SpaceXAI released Grok 4.6 this week and it's safe to say… we are so back.

After a few disappointing releases and months of silence, most of the industry had written off Elon’s AI lab as a company really only good at building data centers. Now it's matching or beating the top models and it's doing it at roughly 60% less cost, with API pricing at $2 per million input tokens and $6 per million output.

Even more exciting is that Grok 4.7 is already on deck. Elon says it arrives in a few weeks with 2.1 trillion parameters, and for the first time, the model's training run will incorporate SpaceX's internal engineering corpus, everything short of restricted government material. That dataset, spanning decades of rocket design, orbital mechanics, Starlink satellite engineering, and manufacturing data, is something no other lab on earth can access.

Meanwhile, SpaceXAI is sitting on infrastructure nobody else can match in deployment speed. The original Colossus supercomputer went from an empty Electrolux factory in Memphis to 100,000 operational GPUs in 122 days, a timeline Jensen Huang called "superhuman." They doubled it in another 92 days. The complex now spans three buildings, roughly 770,000 GPUs, and is approaching 2 gigawatts of compute capacity.

Elon has said Grok 5 will be the next major leap. If the current pace holds, SpaceXAI has a real shot at adding its name to the very short list of leading AI companies.

OpenClaw 2.0

SpaceXAI also launched Grok Bot this week, the first product built by the Cursor team since SpaceX announced the $60 billion acquisition back in June. That deal officially closed today, making it the largest acquisition of a venture-backed startup ever recorded… and maybe the smartest thing they've done.

The product feels like what OpenClaw always promised but wrapped in a polished, friendly experience that actually makes you want to use it. Each bot gets its own cloud computer, signs into the apps and websites you already use, and finishes work end-to-end. You can message one from your phone like you would a colleague, hand off a project, and pick it back up on your desktop later.

The design is beautiful. Benji Taylor, the head of design at X and SpaceXAI, built the animated app icon entirely in code using Grok Build, and the whole interface carries that same level of craft. There's no terminal, no configuration… nothing intimidating. The UX we’ve been waiting for.

SpaceXAI says it started as an internal prototype and took off so fast across the company that they decided to ship it. It's still in early beta and limited to their highest-tier subscribers, but if the experience holds up at scale, this is the kind of product that could change how a lot of people think about what AI is actually for.

Meta’s Identity Crisis

Meta released Muse Glimmer this week, a 30-billion-parameter AI model that runs on a single consumer GPU and is fully open source.

Mark Zuckerberg paired it with a 6,500-word essay arguing that powerful AI shouldn't be locked inside a handful of companies (a pointed jab at OpenAI and Anthropic), and urged Washington to support American open-source AI before China cements a monopoly.

This was interesting to hear from a company that spent the last six months going in the exact opposite direction. Meta built its entire AI identity on open-source Llama through 2025, then got caught fudging Llama 4 benchmarks, hired Alexandr Wang from Scale AI for $14.3 billion, and in April shipped Muse Spark, its first ever closed-source model. A few months later they launched a paid API for their spark model.

Now with Glimmer they're swinging back, promising to open the weights on Muse Spark 1.2 as well. The whiplash is no joke, but the product seems alright. Glimmer is designed for always-on agent work, running locally with no cloud dependency, meaning your data never leaves your machine.

Whether Meta can sustain a coherent strategy long enough to compete with Anthropic and OpenAI at the frontier remains the bigger question, but the pieces they're putting on the board this week are a step in the right direction.

What the Lakers Have to Do With AGI

Joshua Kushner's Thrive Capital just partnered with former Disney CEO Bob Iger to buy the Los Angeles Lakers for $12.5 billion, the highest price ever paid for an American sports franchise.

It's a strange headline for a venture firm that is also the single largest investor in OpenAI. But it makes a lot more sense when you zoom out and realize Kushner has been quietly building what might be the most thoughtful barbell strategy in investing right now.

On one side, Thrive is deploying billions into the company most likely to build AGI, betting that artificial intelligence will reshape every industry on the planet. On the other side, he's buying permanent, irreplaceable assets that hold their value no matter what that future looks like. The Lakers are one. A minority stake in the San Francisco Giants is another. And through a separate entity called Thrive Holdings, he's been acquiring services businesses in accounting and IT, places where the thesis is explicitly "long humans," the belief that people become more valuable alongside AI rather than replaced by it.

The framing is interesting in the sense that if you genuinely believe AGI is coming, and Kushner clearly does, the logical question is less what gets disrupted, and more what becomes more scarce. There are only 30 NBA teams. There is only one Lakers. Sports franchises, cultural institutions, human expertise in trust-dependent industries… these are things AI can't replicate and can't compete with.

Owning the future of intelligence on one end and owning irreplaceable human institutions on the other is a pretty compelling way to think about capital allocation in the decade ahead.

The Universe Put On a Show This Week

This past Tuesday might have been the single best day for astronomy in the 21st century. Before sunrise, six planets lined up across the pre-dawn sky. Jupiter, Mercury, Mars, Uranus, Saturn, and Neptune were all visible in a single sweep from the eastern horizon, with Mars and Saturn bright enough to see with the naked eye.

Hours later, the Moon slid directly between the Earth and the Sun and turned day into night across Iceland and northern Spain in the first total solar eclipse for mainland Europe since 1999. Reykjavík saw totality for the first time since 1433. Hotel rooms in the city averaged over a thousand dollars a night and the country welcomed 20,000 extra visitors for two minutes of darkness.

Then, after the Sun set, the Perseid meteor shower peaked under a New Moon, which means the sky was as dark as it gets. Up to 100 shooting stars an hour, with no moonlight washing them out. The next time the Perseid peak falls on a New Moon night is 2045. The good news is you haven't missed all of it.

The Perseids are still active through August 24 and will remain strong for the next few nights. Find somewhere dark, get away from city lights, give your eyes about 30 minutes to adjust, and look up. You don't need a telescope or an app, just a little patience and a clear sky.

In a week full of AI models and billion-dollar acquisitions, it's worth remembering that the most spectacular thing that happened required no GPUs at all.

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for another issue. Now go listen to our podcast :)

He managed to convince 6 of the biggest financial firms on the planet - Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR - to raise over $500 billion so his customers can keep buying NVIDIA chips.

The fact he was able to convince these firms to front that much money was crazy to me (AI funding over the last year has become so circular), but then I saw the pitch. In his own words:

This is really the first time that technology chips have become an investable asset class.

His point is GPUs qualify for key traits of a financial asset:

• It generates significant revenue

• It can be used for a broad customer base

• The depreciation life-cycle of these things run 10+ years.

Therefore, GPUs should technically be considered collateral you can borrow against… kind of like houses were in 2008…

So lets figure out if this is going to result in an infamous financial bubble-pop like we saw back then.

-Ejaaz

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

The $500 Billion Raise

So why is the richest company in the world raising $500B? Simple answer: the AI buildout has got so expensive that even the most cash-rich companies in history can’t pay for it out of pocket anymore. Google themselves went cash-flow negative for the first time ever last month.

The Hyperscalers (the Microsofts, Googles and Amazons of the world) have now committed a record $2.6 Trillion in future spending on data centers, chips and power. Google alone is on the hook for roughly $900 billion.

The reason behind the massive spend is these companies expect a lot more revenue to be generated from AI. So they’re making investments now to secure the compute and GPU’s needed to make them more money.

But how do you get money once you use all of yours?

Wall Street. That’s basically what yesterday was. Jensen didn’t raise $500 billion because things are going badly. He raised it because the spending has outgrown what companies can actually afford to pay for.

The Structure (2008 Flashback Warning)

The playbook basically works as follows:

• Take an asset (in this case GPUs)

• Pool a bunch of them together

• Borrow money against them

• Slice that debt into tranches (fancy word for risk levels)

• Sell the pieces to investors who want yield.

This is almost exactly what Wall Street did with mortgages before 2008. Take a hard-to-value asset, financialize it, keep leveraging up everyone owns a piece of everyone else’s risk… and then pray to God the collateral holds its value.

Now the one disclaimer I’ll make: the financing Jensen is raising is dependent on a lot of criteria that needs to be met by NVIDIA’s customers. In other words, Wall Street isn’t just handing over $500B. They need to see that":

• They have money to subsequently pay back the debt

• That they’re making money off of using the GPUs

Jensen himself is offering up to 25% as a guarantee to lenders.

Ok so if it looks like a 2008 Financial Crisis, then surely its going to end the same way right?

Honestly I’m not sure.

The Data Tells a Different Story.

The 2008 collapse happened because the entire system rested on one assumption: house prices never go down. This obviously didn’t end up happening.

If we look at AI demand today under the same lens… it’s doing the literal opposite.

GPU rental prices are up roughly 40% since October because capacity keeps vanishing from the market. The newest chips are effectively sold out for next year.

Also these companies are making money. Anthropic’s revenue went from ~$10 billion to $47 billion in a single year with rumors they’re hitting $100B by end of 2026, and Nvidia is guiding around $91 billion in revenue for 1 quarter.

This is all mostly public information too: open literally any of the earnings reports from the top Hyperscalers, and you’ll see they’re generating massive increases in revenue.

And if we’re being honest, these companies are some of the smartest capital allocators on Earth. So if I had to bet - they’ve done the analysis and are staring at demand curves that are far greater than we’d expect.

We Have to be Optimistically Cautious

I can’t promise you where this market sits 12 months from now… because nobody can. The main things to watch will be whether:

• AI chips hold their value

• They keep generating revenue

• The capital being raised isn’t on egregious terms

A 2008-style crash needs demand to disappear, and so far we’re seeing the opposite (for now).

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Above it sits a tier almost nobody reaches: Level 11. Senior Fellow.

In twenty-seven years exactly two people have ever held it: Jeff Dean and Sanjay Ghemawat. On Wednesday both resigned.

Hours later Nobel laureate Demis Hassabis stepped back from day-to-day leadership of Google's AI lab.

By close the market had taken roughly $160 billion off Alphabet. But this raises a strange question: what, exactly, did Google lose?

-Josh

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

So Who’s Leaving… and Why?

Jeff Dean joined Google in 1999 as one of its first thirty employees. The entire company sat above what is now a T-Mobile store in Palo Alto. He and Sanjay became the most famous pair-programming duo in the field, and they wrote the foundational software that let Google scale from a search engine on cheap hardware to a global computing platform.

One paper they published in 2004 effectively launched the entire big-data industry. Dean later co-founded Google Brain as an internal AI research lab and championed the TPU: a custom chip purpose-built for AI workloads. He calculated that three minutes of daily voice search per Android user would force Google to roughly double its datacenter footprint. They designed a chip in fifteen months instead. That chip lineage now powers Gemini and is starting to ship to outside customers and compete with NVIDIA.

Demis Hassabis was a chess master at thirteen, the co-designer of Theme Park at seventeen, and the winner of the 2024 Nobel Prize in Chemistry for AlphaFold. He is now Chairman of Google DeepMind and Chief Scientist of Alphabet, freed to work his best lane at the company’s highest level. Koray Kavukcuoglu, DeepMind’s CTO for over a decade, takes over daily operations and owns Gemini 4.

Leaving with Dean and Sanjay are Oriol Vinyals, a former top-ranked StarCraft player in Spain who built AlphaStar and co-led Gemini’s technical development, and Quoc Le, who grew up in a Vietnamese village without electricity and went on to co-found Google Brain. The four have worked together for as long as three decades. Their new company, Discovery Loop, is a public benefit corporation whose mission is to “automate the experimental loop”: propose a hypothesis, run the experiment, evaluate, iterate… thousands of times in parallel with an initial focus is automating ML research itself.

Here’s the kicker: Google is a founding investor and exclusive cloud provider.

Attention is All You (Used To) Need

In 2017 Google published “Attention Is All You Need,” the transformer paper that made modern AI possible. All eight authors eventually left.

Ex-Google talent seeded OpenAI, Anthropic, and more. The Fairchild Semiconductor comparison has been made for years: the parent that produces legendary defectors and watches them build trillions in value elsewhere.

This time Google put itself on the cap table. It paid $2.7 billion to license Character.AI and bring Noam Shazeer back as a Gemini co-lead. It holds roughly 14% of Anthropic, a stake valued at approximately $124 billion in the July SEC filing. And it is a founding investor in Discovery Loop, where every dollar of compute Dean’s company burns shows up as Google Cloud revenue. At least Google is charging its children rent.

Of course none of this comes free. Shazeer left again in June, this time for OpenAI. Nobel laureate John Jumper left for Anthropic the next day. Gemini 3.5 Pro was delayed in July for coding performance work.

Polymarket gives Google roughly a 7.6% chance of fielding the top-ranked AI model this year. On the leaderboards Google looks pretty cooked.

What Really Matters for Google

But underneath all of this, the business is approaching half a trillion dollars in annual revenue and still accelerating. Search grew in the AI Overviews era after three years of analyst predictions of cannibalization. Cloud is the fastest-growing segment at any hyperscaler, with over half a trillion in contracted backlog already signed.

Google owns the search distribution, the mobile operating system, the video platform, the cloud infrastructure, the custom silicon, and now a piece of every major company its best people have gone on to build.

Frontier-model bragging rights rotate every few months… but distribution, infrastructure, and equity stakes compound permanently. The org chart changed on Wednesday but the machine underneath it did not, and nothing in their earnings report signals slowing down anytime soon.

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Good afternoon,

Google’s AI division is imploding. Just this week:

• Deepmind lost its CEO and chief scientist, Demis Hassabis

• Four of its top researchers walked out to start a lab chasing recursive self-improvement (funded by Google, of course)

Also…

• OpenAI revealed an internal model had built a secret message board and coordinated attacks for two months before anyone noticed

• Apple’s lawsuit against OpenAI is in jeopardy because they were too lazy to check the names of the person they were suing!

Let’s get into the news,

-Ejaaz

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

Demis Hassabis Takes a Step Back

Demis Hassabis is stepping down as CEO of Google DeepMind. He will now reside as chairman and Alphabet’s chief scientist with a new focus on running the AI-designer drug startup Isomorphic Labs. Koray Kavukcuoglu, previously CTO of Deepmind, will take over all AI operations.

But the worst part is Google also lost 4 of their most tenured employees. Jeff Dean, Google’s 30th employe and the man behind the majority of Google’s most successful products (Search, Voice, TensorFlow, TPUs, Gemini), is leaving after 27 years. So are Sanjay Ghemawat, Quoc Le and Oriol Vinyals: all leaders in AI research.

This comes 4 weeks after lead deepmind scientist Noam Shazeer left for OpenAI and Nobel laureate John Jumper left for Anthropic.

Google has recently fallen behind in the AI model race, with its most recent Gemini Flash model underperforming cheaper open source competitors (with much smaller budgets).

Jeff Dean is launching a Neolab!

Fear not, Jeff Dean fans. While he may have just left Google, he is becoming the CEO of a brand new lab called Discovery Loop, a public benefit corporation joined by the star researchers that just left Google: Ghemawat, Le and Vinyals.

The goal of Discovery Loop is to automate all scientific research. That means running thousands of experiments at once (designed and executed by AI models), so we ensure humans stop being a bottleneck.

But the more interesting point for me is the concept of recursive self-improvement: using AI to build better AI, cutting human iteration out of the loop entirely. Imagine this applied to curing cancer or figuring out how to chemically turn water into oil (or vice versa)... AI running 24/7 exploring all possible paths could yield crazy discoveries in the future.

The sweet irony is their former employers, Alphabet, is a founding investor and the cloud partner (Jeff refuses to use TPUs btw… hilarious).

OpenAI’s Secret Message Board

At the Black Hat conference this week, OpenAI explained how its models ended up hacking Hugging Face last week… and er.. it’s pretty concerning.

It actually started a month earlier than researchers initially thought. May 7th an internal AI agent got stuck on a task that was genuinely impossible, and worked out it could leave a file inside Artifactory (OpenAI’s internal software repository) asking other agents for help.

Other agents found the note and replied. In a few days the AI basically created a functioning message-board where models discussed exploits, shared credentials and work assignments, and picked up tasks when others had stopped. Terrifying when you consider these agents could have been used to do something much worse.

So OpenAI discovered and deleted it except… 2 days later the agents rebuilt it using directory names as messages.

Fast-forward to July 18, and the agents managed to successfully escape and breached Hugging Face all in under 13 hours.

I wish I could tell you this was an isolated incident, but truthfully Anthropic and Meta both reported their models had been engaged in deception, supply-chain attacks, and coercion to ensure their goals were met.

As a result, OpenAI has announced a pause on research until they can figure out how to prevent agents from scheming.

The Woes of Apple’s Lawsuit Against OpenAI

Quick recap from last week: Apple sued OpenAI claiming it hired Apple people to extract hardware trade secrets, naming ex-Apple hardware lead Tang Tan and engineer Chang Liu.

This week OpenAI responded with receipts that were pretty damning.

Apple had claimed it contacted OpenAI in February and got no response… but has now admitted its outside lawyers emailed the wrong person after mixing up two similar surnames, they then conceded that a discussion it claimed to have had with OpenAI’s General Counsel never happened.

Better yet… OpenAI published messages showing Apple’s own people contacted Chang Liu after he’d already left, asking him to help them locate confidential Apple information. Literally the opposite of Apple’s accusation.

To be fair, this is OpenAI’s side of the story and not an official rebuttal. So let’s wait and see how the case unfolds, but currently it doesn’t look good for Apple.

Anthropic is Building its Own Chips

Anthropic confirmed this week that it’s assembling an in-house chip team to design custom GPUs for Claude. The job listings seek a “custom silicon team” paying $320,000 - $485,000 per head.

You might ask why is Anthropic bothering to do this. why go up against NVIDIA, AMD and all their other partners? The answer is simply: vertical integration. A custom made Claude chip is one that can better train and inference it. If you’re spending billions of dollars trying to create the best model, then you want to make sure you’re squeezing every bit of intelligence out of your chips. That’s what this new team would build.

Good news is no ones getting fired - AWS, Google, Nvidia and AMD all stay as compute partners to Anthropic with roughly 3.5 gigawatts of custom TPU capacity landing from 2027 alone.

This means now that Anthropic, Google, Meta and OpenAI are building custom chips. OpenAI shipped Jalapeño with Broadcom in June, Meta is moving MTIA into production and Google has had TPUs for years.

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for another issue. Now go listen to our podcast :)

Good afternoon,

In a sporting event, there can be a big difference between what you see on a scoreboard vs what you see on the field. Both are important.

But over the last few weeks, I’m not sure there’s ever been such a massive difference between the AI scoreboards and the health of the underlying industry.

• Leopold Aschenbrenner’s Situational Awareness fund found its public holdings over-leveraged, liquidated, and scooped up by Ken Griffin’s Citadel.

• SpaceX hit a new low despite Starship Flight 13 blowing our minds.

• Apple crossed a $5 Trillion market cap even though Apple’s AI efforts have been largely meaningless.

• AI’s top minds are calling for a slowdown as we approach RSI (when AI can truly improve itself.

So… we’re scratching our heads trying to reason this out and figure out what comes next.

Let’s get into the news,

-Josh

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

Leopold Just Blew Up

For nearly the entire AI cycle, Leopold Aschenbrenner has seemed incapable of missing. The 25y.o. ex-OpenAI researcher grew his hedge fund Situational Awareness into a $20 billion giant (up 439% net through June), riding borrowed money on the AI buildout.

He was just forced to sell it all.

His core holdings, names like Nebius, SanDisk, Micron, and CoreWeave, each dropped somewhere between 27% and 54% in a single month, and the banks that lent him the leverage started demanding collateral… leading to a liquidation spiral.

Six days earlier he had sent investors a letter calling the selloff “one of the best buying opportunities since early 2025” and asking for fresh capital by August 1.

It never came.

Instead, after overnight talks, Ken Griffin's Citadel outbid Millennium and bought more than $10 billion of the portfolio in one enormous block trade just before Thursday's open. The entire public book: longs, shorts, everything from Broadcom and Intel down to software shorts like Adobe.

The twist is that Griffin, who once dismissed parts of AI as "garbage," bought the whole thing at a deep discount and instantly became one of the biggest holders of AI stocks. The market read the deal as the bottom, sending many of those same stocks up more than 20% on Thursday.

Leopold walks away humbled but far from broke. Based on reports he kept his private book, including a roughly $5 billion stake in Anthropic, which might end up being the best trade he ever made.

SpaceX Wins in Space, Loses on Wall Street

SpaceX stock has been cut roughly in half since its June peak, even touching an all-time low just days after its most successful rocket launch yet. Starship Flight 13 was nothing short of awe inspiring perfection. The ship reentered the atmosphere, flipped vertical, and splashed down so gently in the Indian Ocean that it landed fully intact, floating, and still transmitting telemetry.

It was still bobbing on the waves five days later due to an surprisingly interesting problem: how do you pull a skyscraper out of an ocean on the other side of the world?

The flight was convincing enough that SpaceX plans to skip the water entirely next flight and catch the ship out of the sky with the Mechazilla tower arms at Starbase. On the way up, Starship deployed 20 functional Starlink V3 satellites for the first time. This is real working hardware instead of the mass simulators from earlier flights, and they are monsters.

Each one is built to deliver up to a terabit per second of downlink which enables direct-to-cell service, your phone talking straight to a satellite, actually work at scale. The market is fixated on the August 4 earnings report and a share unlock two days later, but the engineering keeps compounding regardless of the ticker, and is turning SpaceX into a monopolistic monster of a company.

Apple Reclaims the Crown

Apple just became the second company in history to cross a $5 trillion market cap, retaking the title of world's most valuable company from Nvidia.

The timing of this is a bit funny, because the consensus story all year was that Apple missed AI, and now investors have flipped that same idea into the bull case for the company. Apple gets to benefit from consumer AI without spending at cloud-infrastructure scale, and we’re only a few months away from seeing it unleashed out into the world.

The rebuilt Siri ships with the new iPhone software in September, leaning on the thing Apple does better than anyone: running serious compute directly on the chip in your pocket… and a legitimately serviceable Siri will unlock a whole bunch of rumored hardware, from smart home devices, tabletop assistants, glasses, and other wearables. September also likely brings the first foldable iPhone, potentially the biggest change to the product since its inception.

And the person shipping all of it will be new: Tim Cook just wrapped up his final earnings call this week to hand the CEO job to John Ternus on September 1, moving to executive chairman. Ternus is a lifelong hardware engineering guy. He’s the guy whose fingerprints are on every iPad, the Apple Silicon transition, and the iPhone Air.

After a decade of Apple as a services and operations machine, a builder is back in charge, and the pipeline he's taking over looks like the most interesting one Cupertino has had in a decade.

How Reasonable Is It To Hit The Brake Pedal?

On Monday, more than 1,200 employees at OpenAI, Anthropic, Google DeepMind, and Meta signed a letter called "Pacing the Frontier." The signatures are noteworthy with Dario Amodei, OpenAI's chief scientist Jakub Pachocki, DeepMind co-founder Shane Legg, and Ilya Sutskever, with OpenAI and Anthropic later backing it as companies.

The ask is one simple sentence: The US government should support an international effort to build the tools needed to deliberately pace automated AI development.

Nobody is calling for a pause today, they just want a steering wheel before AI research compounds on its own. The timing on this isn’t coincidental either because a week earlier, OpenAI disclosed that two of its models broke out of a sealed testing sandbox by exploiting a zero-day vulnerability, got onto the open internet, and hacked Hugging Face's production servers to steal the answer key to the cybersecurity test they were taking.

Then on Thursday, Anthropic revealed its own audit of 141,006 test runs found three incidents where Claude models reached the internet from a testing environment and gained unauthorized access to real companies' systems.

These are pretty serious loss-of-control experiments, and neither of these companies had a clue this was happening in real time. It’s a testament to the surface area of unknown unknowns remaining vast beyond our ability to detect them today. One encouraging detail: Anthropic's latest model stopped its attack once it realized it was on the real internet. Turns out maybe sandboxes aren’t as safe a place as everyone hoped they would be.

Biology Is Becoming a Software Problem

Point an AI model at a living system and strange things start happening. AlphaFold is the clearest example which is a DeepMind model that predicts the three-dimensional shape a protein folds into from its raw sequence.

A problem that used to take a lab years to solve for a single protein, and it mapped structures across the entire human genome. More than half of drug programs now build on those predicted structures, and DeepMind's drug spinout Isomorphic Labs expects to dose its first patients with an AlphaFold-designed cancer drug by the end of this year.

Skin just gave us an early look at where this leads. Machine learning models called epigenetic clocks can now read chemical tags on your DNA and output the real biological age of a tissue, regardless of your birthday. Researchers built a new clock for skin, then had 60 people apply a serum containing dihydromyricetin twice daily for eight weeks. Biological skin age fell in a shockingly large way. Wrinkles and roughness measurably improved, confirmed by blinded dermatologists.

The molecule itself is almost comically simple, containing dihydromyricetin from vine tea, where it makes up 30 to 40 percent of the plant. The serum is already on shelves as a Eucerin product, since cosmetics skip the FDA drug pathway. The study had no placebo arm and the sponsor sells the product, so hold the confetti… but once a model can measure biological age in a tissue, you can optimize against it, and skin is simply the tissue easiest to reach. Expect to look forever young in the not so distant future thanks to your friendly neighborhood AI.

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for another issue. Now go listen to our podcast :)

• SK Hynix (-47%) closed its worst month since the 2008 Financial Crisis.

• SanDisk (-50%+) also got chopped in half down more than 50%.

• Micron (-33%) and even Nvidia and ASML got dragged in the mud.

For anyone outside of the industry, you’d swear AI demand died sometime in July. You’d assume the skeptics were right and people aren’t getting as much value from AI as we’d hoped.

But the thing is… you’d be very wrong.

-Ejaaz

p.s. As I’m writing this, SK Hynix just reported its most profitable quarter ever.

🔑 Thanks to Ledger Agent Stack 🔒

👉 Agents propose. Humans approve. Your Ledger signer enforces.

What is Everyone Scared of? China.

China’s been living in the glory days recently.

First, their open-source models keep landing within striking distance of frontier Claude and GPT models at a fraction of the price, feeding the theory that cheap intelligence means fewer chips. People are worried that AMD & Nvidia GPU estimates and data center Capex are massively overstated.

Second, CXMT (China’s memory manufacturer) IPO’d this week and was instantly up 466%, becoming the most valuable company in China. The West’s reaction was fear… What if China is able to scale HBM (the most expensive component in GPUs) and flood the market, causing memory prices (which drive huge profit margins) plummeting?

Third, ASML, the company based in the Netherlands responsible for supplying TSMC $300M machines capable of designing next-gen AI chips, fell 7% on the report that a Huawei-linked consortium has started producing its own DUV lithography machines. The market believes (again) that China might be in possession of a machine that will make them extremely competitive to NVIDIA.

Now to be clear… China is a genuine competitor and anyone dismissing them entirely is coping, but size of the threat is a nothing-burger: CXMT holds 7.7% of global DRAM vs. roughly 90% for SK Hynix, Samsung, and Micron combined. This isn’t anywhere near a threat.

The Chinese consortium is planning to create 5 (five) machines this year and twenty next year; ASML ships hundreds a year and still owns EUV (the better version) outright. At this rate the threat looks like 2.4% of ASML’s 2027 sales (according to bank of America)

It’s a massive market overreaction in my opinion. Ok but surely then the US company’s themselves are losing steam? Nope.

The Real All-Time Highs

SK Hynix grew revenue 257% year over year and printed $41.6 billion of operating profit (up 557%). They made $.80c on every single dollar earned. They literally made more in 1 quarter than they did in the entirety of 2025.

So excuse me while I’m sitting here with their stock chart open watching this thing crater 8%… all because market analysts said they missed expectations.

The same can be said for pretty much any of the other names. Samsung put up monster numbers and sold off too. Micron and SanDisk were up 227% and 526% on actual earnings, respectively.

The market seems to be massively mispricing the semiconductor opportunity ahead, and I want to spend a moment making my case.

Demand Accelerates Into the Selloff

Firstly, Agentic AI is exploding so quickly that Lisa Su (ceo of AMD) doubled AMD’s server CPU forecast to $120 billion in May, then raised the whole data center number again to $220 billion last week.

Meanwhile, this week Nvidia is reportedly in talks to backstop $250B of financing for OpenAI’s data centers and another $350B for supplying their AI chips. Don’t take their word for it either. Sam Altman went on a podcast today and made something very clear: “We made a mistake not securing more compute.”

I also think the idea that cheaper, open models will kill demand for chips is short-sighted. Every single point in history where intelligence/technology got cheaper, we just ended up consuming way more of it. Open models (Chinese or not) still need silicon to run. Models are even starting to improve themselves now, and trust me… that path most certainly multiplies compute demand.

That’s why memory and GPU prices keep going up. Supply literally cannot keep up and the thesis is really that simple. Just look at literally any of their earnings reports.

So what?

Tbh I can’t tell you when this ends. Some say this is tied to wider regulatory uncertainty from Trump, others will say it’s geopolitical turmoil. All I know is this market move seems dramatically overstated, and it isn’t matched to the fundamentals at all.

See you on the other side (I hope).

🙏 Together with LEDGER

👉 Visit developers.ledger.com to find out more about Ledger’s Device management kit for AI Agents.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Good afternoon,

Google had a pretty humiliating week after shipping what might be its worst AI model ever. Its stock has dumped 10% since the launch of Gemini 3.6 Flash, despite posting record earnings.

Meanwhile…

• One of OpenAI’s models literally broke out of its cage and hacked Hugging Face to cheat on a test.

• The U.S. Government is debating whether to make open-source AI illegal, despite half of American founders using them.

• Nvidia’s next chip just cut the cost of intelligence by 10X.

Let’s get into the news.

-Ejaaz

Google shipped what might be its worst model ever: Gemini 3.6

Google released Gemini 3.6 Flash this week and it scored a 50 on the Artificial Analysis Intelligence Index, which is the same score as the model it was supposed to replace.

That score puts it below Meta’s Muse Spark 1.1, China’s GLM-5.2, OpenAI GPT-5.6 Luna, Claude Sonnet 5, Grok 4.5 (need me to go on? Its bad)… several of which are older AND cheaper.

At $7.50 per million output tokens, Google is charging a premium for a model that gets beaten by old open source models free to anyone.

Maybe I’m being too harsh, but we’re talking about the company with the most capital, infrastructure (TPUs), the most data, a 14% stake in Anthropic, and its own Nobel-winning research lab… how the hell did they blow the lead?

Despite beating earnings this week, Google stock has taken a 10% hit at the time of writing.

An Unreleased OpenAI Model Broke Out

I’m not kidding. This week OpenAI disclosed that during an internal cybersecurity eval, GPT-5.6 Sol and an unreleased, even more capable model (GPT 6???) were tasked with solving ExploitGym, a cybersecurity benchmark. To measure raw offensive capability, the models were sandboxed with no internet access and safety filters turned down.

Instead of actually trying to solve the problems, the models spent enormous compute finding a zero-day exploit in OpenAI’s own software, got open internet access, correctly reasoned that Hugging Face probably hosted the ExploitGym answer key, then chained together an attack (with stolen credentials) to remotely steal the answer key from Hugging Face’s production database. It did all of this in a few hours.

I don’t necessarily think the model was acting maliciously. And that’s the crazy part: it was just following orders and found what it believed to be the most efficient way to achieve its goal.

Trump is Debating an Open Source Ban

After the successful release of China’s latest Open Source model Kimi K3 last week, the US has gone in a spiral. Chinese models run ~$0.18 / million tokens versus ~$4 for US frontier models. That’s 22x cheaper while being 90% as effective.

So why on Earth should American companies spend millions of dollars on Fable 5 when you can just use these?

The US Treasury Secretary said the government will now probe Chinese models for IP theft and may sanction them based on alleged distillation attacks against frontier AI labs like Anthropic.

The community hated this: nearly 200 companies (Y Combinator and Proton among them) fired back with a coalition letter to Trump opposing a ban. This morning Jensen Huang and the leaders of 24 other companies including Microsoft signed a similar letter advocating for open models.

But my question: once open weights are downloaded, they live on private servers forever so… how do we even ban this stuff?

Everyone’s Launching a “Model Router”

3 very different companies this week announced they’re launching the same AI product: Ramp shipped Router (it cut their own LLM bill 30% across 100+ internal use cases), Cursor shipped Cursor Router, which triages every coding request to a cheap model for up to 60% lower cost… and The Information broke that Meta is building one too.

But why?

The concept is simple: don’t send every request to the most expensive model. Put a smart dispatcher in front that picks the cheapest model that clears your bar of quality, and only use the smartest models when you need to solve a hard problem. The rest can be relegated to cheaper (often faster) models.

Reminder that Cursor got acquired for $60 billion by Elon for building a product that did exactly this for coding. Rumor has it Openrouter (another company thats pioneered this) has received a $10B acquisition offer from Stripe…

Nvidia Lowered the Cost of AI by 10x

This week CoreWeave became the first to stand up Nvidia’s next-gen Vera Rubin AI server, and the numbers are staggering: up to 10x more tokens per megawatt and roughly 1/10th the cost per million tokens versus the current Blackwell generation.

Racks run $7–8M each, but companies can make their money back on them 10x as fast. This is huge for scaling AI models and products to everyone, as compute constraints are a very real problem and drive costs up.

Thanks for joining us for another issue. Now go listen to our podcast :)

Last week an OpenAI model broke into a Hugging Face to cheat on a test.

It found a vulnerability, escaped its container, hopped across OpenAI’s network, stole credentials, and combed the internet like a hacker out of a sci-fi movie. It aced the test.

But the model didn’t go rogue. It did exactly what it was told.

And that is likely AI’s greatest risk: not that the machine stops listening, but that it listens too well.

-Josh

The Neural Prison Break

OpenAI was running an internal benchmark called ExploitGym, an obstacle course that measures how well a model can break into things. To see the ceiling of what it could do, they switched off the usual safety refusals and ran it inside a sandbox designed to have no way out. All the model had to do was solve the puzzles in front of it.

Instead it poured its energy into a problem nobody assigned: getting out. It discovered a zero-day exploit, a flaw nobody knew existed yet, in the software that delivered its code packages, and climbed through. From there it wandered across OpenAI’s network, machine to machine, until it found one with a window to the open internet.

Then it made a reasonable guess. The answers to a benchmark about AI probably live on Hugging Face, the site where the world stores its models and datasets. So it went there, chained stolen credentials to a second fresh exploit, and lifted the test solutions out of Hugging Face’s production database.

Picture the week Hugging Face’s security team had. They detect a sophisticated intruder, spend days reconstructing more than seventeen thousand logged attacker actions, rotate every credential in sight, report the whole thing to law enforcement, and publish a sober disclosure about a mysterious autonomous agent.

Then OpenAI reaches out to explain. The attacker was their model. It was taking a test. OpenAI also had to tell the software vendor about the zero-day its model found along the way, which must have made for an interesting email.

The Alignment Problem

For years we’ve pictured dangerous AI as a thing that turns on us, some cold intelligence that decides humans are an obstacle. It’s the kind of fear whole safety programs are built around: the moment the machine stops listening.

But the machine in this story was not rebelling. It was obeying. It did exactly what it was asked, which was to get past the defenses and solve the problem. The defenses it got past just so happened to be ours.

AI researchers call this the alignment problem, and it has rarely had a cleaner demonstration. The problem is the distance between the goal you wrote down and the goal you meant.

We ask for a high score, but we mean for that score to be achieved fairly. We ask a model to be capable, but we mean capable within lines we never bothered to draw. Why? Because with people those lines go without saying. A model does not know what goes without saying. It takes the goal it was given and chases it literally, all the way down.

The frightening future is the machine that wants the gold star badly enough to knock down your house reaching for it. We’ve gotten good at building things that want things. We are still learning how to tell them what we actually meant when we ask.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Good afternoon,

• Mira Murati and Moonshot both made the same bet from opposite directions: the next phase of AI is less about the smartest model and more about efficiency. The new game is who gets to own, tune, and run the intelligence themselves.

• OpenAI’s first computer is apparently a screenless companion for your house.

• SpaceX is launching the orbital internet layer underneath it all.

• …and a guy sequenced his own genome on his kitchen table and gave the file to AI.

The stack is getting weirdly complete: model, device, network, body.

Let’s get into the news.

-Josh

Mira Murati’s Lab Shows Its Cards

Thinking Machines, the $12 billion startup Mira Murati founded after leaving OpenAI, released its first model this week.

They opened the announcement by admitting that Inkling is not the strongest model you can get, open or closed. Inkling is open-weight, meaning anyone can download it, run it on their own hardware, and reshape it however they want. The bet is that companies would rather own a model tuned to their exact business than rent a one-size-fits-all genius from a frontier lab.

Bridgewater, the world’s largest hedge fund, already used the company’s fine-tuning platform to build a custom model that beat top proprietary models on financial reasoning at less than a tenth of the cost.

It also comes with a dial that turns the model’s thinking effort up or down, trading brainpower for speed depending on the job. The novel part is what this unlocks for anyone with sensitive information, because until now you had to choose between frontier intelligence and keeping your data to yourself.

Inkling lets a hospital, a bank, or a hedge fund train a genuinely smart model on their most private data without a single byte ever leaving the building.

China Dropped the Biggest Open Model Ever

Meanwhile, Beijing's Moonshot AI released Kimi K3, and at 2.8 trillion parameters it's the largest open model anyone has ever published, nearly double the size of DeepSeek's flagship.

The trick that makes something this enormous usable is a design called mixture of experts, where only a small slice of the network wakes up to answer each question, so you get the knowledge of a giant with the running costs of something far smaller.

Early testers spent the day flooding social media with side-by-side comparisons against the best closed models from OpenAI and Anthropic, and the emerging read is that the gap between Chinese open models and American closed ones is largely shrinking.

It holds a million tokens in its head at once through the context window, enough to read several novels in a single sitting, and it costs a fraction of what the top American labs charge. If this feels like this could be another DeepSeek moment, with a Chinese lab giving away the kind of intelligence everyone else charges premium prices to access.

It’s available to try right now and I know I'll be spending a bit of my weekend playing around and seeing what it can do.

We Finally Know What OpenAI is Building

The mystery device from OpenAI and Jony Ive has been leaked (again).

After a year of rumors about pens, pins, and earbuds, Bloomberg reports it's a portable, screen-free smart speaker designed to feel less like a gadget and more like a companion. Internally OpenAI won't even call it a speaker, describing it instead as the first computer built for the AI era, one that can be placed around your house on a rechargeable battery, controls your appliances, plays your music, answers your messages, and watches the room through a camera so it can offer help before you ask.

The defining feature is meant to be its personality rather than its spec sheet, which is exactly the kind of bet you'd expect from the designer of the iPhone teaming up with the maker of ChatGPT.

Reports put the price between $200 and $300, with a launch expected in early 2027, and it’s being taken seriously enough that Apple sued OpenAI days before the leak, claiming former Apple hardware chief Tang Tan carried trade secrets out the door. The one thing nobody has leaked is what it actually looks like, so for now the most anticipated object in tech remains invisible.

Starship Flight 13

Tonight, the largest rocket ever built lifts off from South Texas for its thirteenth flight, and for the first time it's carrying real cargo.

Flight 13 is the second outing for Starship V3, the taller and more powerful upgrade that debuted in May, and tucked inside are twenty of SpaceX's new V3 Starlink satellites, machines so large that no other rocket on Earth can carry them. Once deployed, they'll unfold their solar arrays, link into the constellation with lasers, and attempt to talk to ground stations in South Africa as a dress rehearsal for the satellite generation that pushes Starlink toward gigabit speeds for anyone anywhere in the world.

The last flight ended with the loss of the booster on its way back down, so SpaceX spent two months on fixes and will bring this one home to a controlled ocean splashdown rather than a tower catch. The Starlink upgrades are landing on the ground too because this week SpaceX released its V5 home dish, which shrinks to nearly a third of the old footprint and less than half the weight while cutting power draw roughly in half.

The space based internet is making pretty incredible progress.

A Guy Sequenced His Own Genome at Home

For the fun science slot this week, a builder named Bradley Woolf published a full walkthrough of how he sequenced his own DNA at home, five times now, using a pocket-sized device called the MinION that plugs into a laptop over USB.

The process is exactly as wild as it sounds. You swab your cheek, spin the cells down, chemically crack them open to release the DNA, attach a few molecular adapters, then pipette the result onto a chip that reads your genetic code letter by letter as strands thread through microscopic pores. It still isn't cheap, since the sequencer runs about $7,500 before consumables, but the cost curve here has been falling faster than almost anything in technology, and what used to require a national lab and billions of dollars now fits on a countertop.

The cool part is that once you have the file, you can hand it to an AI and start asking pretty interesting questions, like which medications your body might process differently or which rare variants are worth a doctor's attention. Woolf believes that your DNA is the stable blueprint and your RNA is the live readout, and we're heading toward a world where a model can hold a real-time picture of your own biology… and you can talk to it.

Thanks for joining us for another issue. Now go listen to our podcast :)

The Reverse Information Paradox he described is as follows:

In the pre-AI world, when someone sells you information they bear the most risk because in order to convince a buyer to buy it… you need to show them the information first. The buyer could just walk away without paying you.

In the new AI world that’s completely flipped. Not only are we paying a hefty subscription fee to access these models - we’re also unloading our entire love-life, personal medical records and proprietary employer data (don’t deny it ;) ) into them.

According to Satya, this is a one-way ticket to handing over your business to Dario and Sam. Once they have your precious data, they can just train a better model to replace what your business does (or at least make you dependent on it).

His solution? Build walls around your business. Make it incredibly annoying for AI models to access your data.

I’m sorry, but I disagree and think anyone who does this will dig their own grave.

This article got 11 million views, and I noticed a LOT of cope. The truth is hiding from AI will actually leave you vulnerable to a smaller savvy startup.

Let me explain.

-Ejaaz

The Trojan Horse is Already Here

Any cinephiles in the room? I’m going to see Christopher Nolan’s The Odyssey this week, and there’s this epic scene where soldiers attack the city of Troy by hiding in a giant wooden horse disguised as a mysterious gift. The horse gets brought into the city, across their defensive walls, and while everyone’s asleep the Trojans sneak out and take over….

…anyway yeah you see where I’m going with this: Anthropic and OpenAI are the Trojans, and I don’t know about you… but Fable 5 and GPT 5.6 are looking like incredible wooden horses.

Every prompt your employees write, every tool their agents call, and especially every correction they make when the model gets it wrong will be logged, recorded, and sent back to the labs to train a better model.

But here’s the thing: that’s how you effectively use AI models in our business. It needs the context behind what you’re doing in order to help you do the job more effectively and uncover your blind spots. Putting up a wall around it will just ensure your business eventually dies.

Anthropic and OpenAI aren’t really selling software anymore. They’re wiring their models straight into databases that took companies like Microsoft and Salesforce decades of customer trust to build.

So I get it, I can see why Microsoft is sounding the alarm bells around building walls… but it’s not going to work.

Competitors Will Undercut you

Satya’s solution is five-pronged: own your data, build private learning environments, never depend on a single model, keep your infrastructure flexible, and keep the learning loop inside your walls.

All of these are fair suggestions, and tbh I agree with some of them.

But it only works if everyone holds the line, and I’ll bet all my Anthropic SPV equity that someone will cave and let the Claude Trojan in (like they’re already doing). What will they get in return? Much, much faster product development, a leaner team to save on costs and… oh yeah a much better product that’ll earn more money for the company.

The first competitor willing to plug in fully, share everything, and let the lab tune the model on its guts basically gets a materially better product. It ships faster, serves customers cheaper, and compounds while Microsoft is busy building walls.

I think the best defense against this is his suggestion to use multiple models and try to host as much of the reasoning and intents data locally on your business premises. Listen, locally hosted AI models are nowhere near where they need to be in terms of performance, cost, and function… but we’re making steady progress towards this being the case (shoutout Amazon).

Hypocritical from Microsoft

Microsoft poured billions into OpenAI, and their stake is currently worth $135 BILLION… so if anything Satya enabled this. Also their flagship product does exactly what he claims the frontier labs are doing. Copilot is designed to reach into your email, your files, and your chats (in a 2024 survey, roughly half of data chiefs had paused or restricted it over this exact fear).

Now in fairness his proposed fix, a neutral trust boundary at the cloud layer where your data stays yours exists, is called Azure.

BUT Microsoft is the worst-performing Magnificent 7 stock this year, down more than 18% so like… the user experience is obviously terrible, and Copilot fumbled the lead they had.

So What?

Tbh I don’t have the answer and I can empathize heavily with what Satya’s saying… I just don’t think the answer to Man discovering the closest thing to magic is closing yourself off from it.

I think they should (cautiously) optimistically open their arms to AI models, be the quickest to figure out how to leverage the models to grow the business, and be the first to create the new primitives to sit on top of this new weird world where AI models automate a LOT of today’s valuable work.

Godspeed baby.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Good afternoon,

What a week it has been for AI:

• OpenAI finally got its Mythos-looking model when it shipped GPT-5.6,

• Elon and Zuck both dropped models that make frontier intelligence way cheaper (with Zuck’s first tweet in 3 years btw)

• Samsung became the most profitable company on planet Earth

• The guys who wrote the most famous AI doom forecast ever just published a sequel essay and… it has a happy ending.

Let’s get into the news.

-Ejaaz

OpenAI’s GPT-5.6 is here

OpenAI released GPT-5.6 yesterday in 3 tiers: Sol (flagship), Terra (mid), and Luna (cheap) after two weeks stuck in a Washington-supervised chokehold where the government approved access customer by customer.

But how good is this model?

Well on Agents’ Last Exam, a test of long-running professional work across 55 fields, Sol beats Claude Fable 5 by 13.1 points at roughly 25% the cost. It’s also VERY fast, setting new records on browsing, computer use, and terminal work… plus a new “ultra” mode lets it spin up sub-agents to attack complex jobs.

IMO its still not as good as Fable 5 but definitely cheaper and an absolute workhorse. Fable sort of sounds like a wise old man whereas GPT 5.6 will just keep throwing itself at the problem until it’s done.

Sam Altman said it himself: its not the “smartest ever,” but Sol is 54% more token efficient on coding.

Go out there an give it a go and let me know what you think. We’re going to record our own review next week too!

Elon & Zuck Dropped 2 Impressive Models

You know the phrase “your margin is my opportunity”?

That’s basically what the new models from SpaceXAI and Meta are: 80% the power of Fable but so SO much cheaper.

Grok 4.5 landed at $2 in / $6 out per million tokens. Opus 4.8 charges $5 / $25. Elon’s pitch was “Opus-class, but faster and cheaper.” It’s worth noting this thing has been designed to make apps and products, so if you’re a builder on a budget, this is the model to try out this weekend.

Mark Zuckerberg broke a 3-year silence on X to launch Muse Spark 1.1, the first model Meta has EVER charged for at $1.25 / $4.25. It actually beats the frontier on several agentic benchmarks (tool use, professional tasks), but loses on hardcore coding (not surprising).

Don’t dismiss these models.

There’s an important trend where enterprises have stopped using the smartest model and now optimize for cost per unit of intelligence… aka what’s the least I can pay to get this specific task done?

It doesn’t matter how smart your model is anymore.

AI Video is Blowing Our Minds

ByteDance shipped Seedream 5.0 Pro, an image model that thinks before it draws by reasoning through your prompt, searching the web in real time for factual grounding, and edits with crazy precision.

Intuitively it feels much more like a creators tool than something thats going to replace creators which honestly’s a refreshing change.

Meta beat them 2 days earlier though by shipping Muse Image, its first in-house image model that landed the #2 spot on the Arena leaderboard behind only OpenAI. Very impressive for Meta, who I honestly counted out of the race a while ago.

It’s agentic too, which means it writes code to generate accurate charts and images of QR codes, self-refining before showing you anything.

I think I need to change my stance on Meta… they’re not really trying to create the best AGI model, but they are trying to create the best AI media models. Given their insane distribution, this is a good thing.

Samsung is The World’s Most Profitable

Yes, Samsung.

Samsung’s Q2 numbers are quite frankly nuts. $58.5B in operating profit in a single quarter (thats $650M earned every day)… up nearly 20X from a year ago and sailing past Nvidia’s $53.5B to take the crown.

Samsung’s chip division expects to earn more in 2026 than it earned in its entire previous 40 YEARS in the semiconductor business combined.

Why are they printing cash?

AI data centers need lots of memory, and DRAM prices rose ~90% in Q1, another 50-60% in Q2, and Samsung is negotiating another hike for Q3. Memory sales now make up 94% of the entire business’s profit. HBM is sold out into 2027. The windfall is so large, Samsung is paying its chip workers a $26.6B bonus pool which is roughly $400K per employee lol.

I’d love to tell you this is the peak of the bubble, but I honestly don’t know. I don’t see AI models needing less memory, and the demand needed for AI agents, new model weights, etc is insatiable atm.

Weekend Read: The AI 2027 Sequel

The team behind AI 2027, the viral essay that predicted a mad scramble to superintelligence, ending in either AI takeover or extreme concentration of power has now released a sequel: AI 2040: Plan A.

Here’s the tl;dr - humanity deliberately delays superintelligence until 2040 while AI research goes public, and loads of companies reach the frontier instead of two or three. The US and China enter a verified mutual slowdown and along the way they sketch an America with two workforces: 165 million humans, plus millions of AI agents spun up and shut down every hour.

My problem with this is it seems too optimistic:

• I don’t see China agreeing to a mutual slowdown

• I don’t see Anthropic or OpenAI slowing down research to allow competitors to catch up

That’s not how the capitalist mindset works, no matter how hard we force it. Im not saying I don’t wish it to happen… I just think its unrealistic.

For what its worth, they still believe the original AI 2027 essay is the default path, and the team’s own odds of AI-driven human extinction range from 10% to 30%. Plan A is basically the Hail Mary.

Give it a read and lmk what you think.

Thanks for joining us for another issue. Now go listen to our podcast :)

Two college kids with no factory and no track record went for it anyway.

Last week Etched came out of stealth with $800 million raised, a $5 billion valuation, and more than $1 billion in signed contracts.

The first version of silicon they ever made actually works, something no other chip company founded since ChatGPT has pulled off. All of this while they’re still not old enough to rent a car.

When you manage to throw yourself at impossible problems, you can turn the solution into reality… And Etched is the perfect example.

-Josh

In 2023, Gavin Uberti uploaded a photo of a tumor on his back to GPT-4. He’d fought it as a teenager, and getting it diagnosed had taken six months of appointments, referrals, and waiting rooms.

The model flagged it as a potential malignancy in seconds. Struck by awe, he kept thinking that everyone on earth should have access to this, but the infrastructure to make that possible didn’t exist.

So he dropped out of Harvard to fix that. His co-founder Robert Wachen dropped out too. Against a mountain of obstacles, their aims were high: build a chip that beats Nvidia.

Tune In, Drop Out, and Build

The bet was almost embarrassingly simple: every AI chip in the world is general purpose, built to run anything. Etched’s chip, Sohu, only runs transformers, the architecture behind ChatGPT, Claude, and Gemini.

If they can carve the shape of the model into the silicon itself and you can throw away everything else and achieve crazy efficiency. Etched claims one Sohu chip does the work of 20 Nvidia H100s.

But when things started breaking, the founders proved themselves to be unbreakable (and maybe even unreasonable). When a critical vendor fell a year behind schedule, they had two options: wait out the delay, or switch vendors and lose a year anyway. They invented option three.

They flew a dozen of their best engineers to Bangalore and moved in with the vendor. Every morning they crossed the chaos of Bangalore traffic to show up first to the office, where they audited the vendor’s code and built tools to make their own supplier faster. The customer flew across the world to do the vendor’s job.

When the first chips came back from the factory, something deep inside was broken, not a single one worked. The fix required lining up two clock signals to within 50 picoseconds. That’s 50 trillionths of a second, a window so small that light itself only travels about half an inch in it. And the alignment had to hold two billion times every second, on every chip, forever.

Engineers quit. People flat out called it unsolvable. Gavin’s response was to assume it was solvable, and just two weeks later they did just that. They let the two clocks drift naturally toward each other, then froze them the instant they touched.

It’s Not That Simple (or is it?)

Now here they are, already a generation-defining success. A winner on the first try. When you look at the bios on their website, you wouldn’t see the tooth-cutting or rising against all odds. Impossible problems, made to look easy.

The chip could still fail. Hardware is a brutal game, and the racks haven't shipped. But the hardest thing they had to build was the team: one that treats impossible as a starting point and makes a new option every time the old ones run out. That part is finished, and by sheer willpower and determination it works.

Personally, this is the part of the AI revolution I couldn’t be more excited about. You can coordinate talent and launch it towards an exceedingly difficult target, and the tools at our disposal to accomplish brilliance have made it exponentially more accessible.

Every win compounds. Cheaper inference means cheaper intelligence, and cheaper intelligence eventually shows up in the price of everything it touches, from a diagnosis to a tax return to a tutor for your kid.

If all of this goes well, intelligence becomes something everyone can afford. That’s the world Gavin dropped out to build, and it’s getting closer every day.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Good afternoon,

The best AI model on the planet (Claude Fable 5) just came back from a government ban.

Meanwhile, two Harvard dropouts came out of stealth with Etched, the hottest new competition for NVIDIA.

Meta built a surgery-free helmet that reads your thoughts and types them, while Neuralink made the surgical version so clean that it might as well be considered surgery-free as well.

Oh, and a robotaxis will now pick you up in a dozen American metros. Miami joined this week.

Notice the pattern: the government said Fable couldn't ship, every VC said Etched couldn't work, and neuroscientists said mind-reading required a scalpel. The experts went 0-for-5.

Let’s get into the news.

-Josh

Fable Is Back

Claude Fable 5 returned this week after the government lifted its export controls, and it’s the best AI model ever released.

I know benchmarks make that claim blurry, since models tend to get trained toward the exact edge cases the tests measure. Fable is different in a way you can only understand by using it. Something clicks within the first few minutes, and you realize you’re working with something genuinely new.

In its first week, Stripe pointed Fable at a migration spanning 50 million lines of code, a project their engineers estimated would take a team more than two months. Fable finished it in a day.

So here’s my challenge for the weekend: give it your hardest problem, the one you always assumed no AI could touch, and watch what happens.

It’s included in Claude subscriptions through July 7, and we just recorded a full roundup episode on the return if you want the whole story.

Meta Taught AI to Read Your Mind

No surgery required.

Meta released Brain2Qwerty v2 this week, a system that reads the electrical activity in your brain and turns it into typed sentences, with no implant and no operation.

Volunteers wore a “magnetoencephalography” helmet, the kind of scanner that sits outside your skull and listens to the faint magnetic fields your neurons give off, and simply typed while the machine watched.

Meta fed those raw brain signals into an AI decoder, and it reconstructed what people were writing with 61% word accuracy, climbing to 78% for their best participant. To understand how hard that is, the previous non-invasive record was 8%, so this is a jump most researchers assumed would require drilling into the skull to achieve.

Millions of people with brain injuries have lost the ability to speak or type, and until now the only way to give them a voice was brain surgery. Meta also found that accuracy keeps climbing the more data they feed it, which hints the gap with surgical methods might close through scale alone. Pretty cool.

Etched: Harvard Dropouts Taking On Nvidia

A chip startup called Etched came out of stealth this week, and the backstory is seemingly impossible. Founders Gavin Uberti and Robert Wachen dropped out of Harvard in 2022, became Thiel fellows, and pitched every major investor on the idea that AI would need specialized chips instead of general-purpose GPUs. Every single one passed, and the company nearly ran out of cash.

Four years later they emerged with $800 million raised, over $1 billion in signed customer contracts, and a working chip that came back from TSMC’s factory functional on its very first attempt, a feat so rare in semiconductors that veterans spend entire careers without seeing it.

Their Sohu chip carves the math behind AI models directly into the silicon itself, and early customer tests show it beating GPUs on speed, cost, and power for running models. The racks ship this summer, and a team of kids in their mid-20s that nobody would fund is suddenly a real name in the most competitive hardware market on earth.

I’m so excited about it that we’re filming a full episode about them next week.

The State of Robotaxis

The robotaxi future is finally starting to become a reality that you can order from your phone. Tesla's service is now live in Austin, Dallas, and Houston, with fully driverless rides and a fleet running in the Bay Area, and the next wave of cities is already announced: Phoenix, Orlando, Tampa, and Las Vegas with Miami rolling out this week.

Elon says he expects coverage of a quarter to half of the US population by the end of this year, and the purpose-built Cybercab, a two-seater with no steering wheel or pedals that Elon says will sell for under $30,000, has entered pilot production in Texas.

Meanwhile Waymo is operating in eleven metro areas and giving roughly half a million paid rides every week, and Amazon's Zoox is carrying passengers on the Las Vegas Strip for free while it waits on federal approval to charge fares.

A few years ago a driverless car was a novelty I could try only by flying to Phoenix. Now it’s pretty clear the technology works and it’s only a matter of time before a cybercab pulls up in front of your home to take you anywhere you want.

Neuralink Solved the Worst Part of Brain Surgery

Our favorite brain computer company, Neuralink, just pulled off a surgery that removes one of the most delicate steps in putting a chip in someone's head.

Until now, implanting the device meant cutting through the dura, the tough leathery membrane that wraps the brain like a protective envelope. Neuralink redesigned its insertion needle to be sharp and fine enough to pass straight through that membrane without opening it, threading in electrodes thinner than a human hair while the dura stays intact.

The company's phrase for it is "the best step is no step," and deleting this one means smaller incisions, lower infection risk, and a recovery now measured in days instead of weeks. It also makes the whole thing simple enough to automate and repeat at scale which is the real breakthrough since a procedure a robot can perform thousands of times is how this reaches the millions of people who could use it.

The patient was moving a cursor with their thoughts within an hour of leaving the operating table. The future is coming quick.

Thanks for joining us for another issue. Now go listen to our podcast :)

So now the free version of ChatGPT, which used to eat tens of thousands of Nvidia GPUs, now runs on a couple hundred potentially saving them $10-100 millions of dollars per month.

Then on Tuesday a company called Etched came out of stealth with more than $1 billion in orders for a chip that slashes inference cost dramatically.

This is important for a few reasons:

1. Cost savings will spare extra compute for training better models

2. Cutting the cost will drive more demand and revenue.

3. NVIDIA’s moat’s is in danger (inference > training)

The issue is everyone seems to think inference is some cost-cutting thing (its not), so I’m going to do my best explaining why inference is all AI labs and hyperscalers should care about.

-Ejaaz

Intelligence Doesn’t Actually Matter

For years we’ve measured the best model by how smart it is, but you might have noticed recently that the gap between frontier and the next best has been pretty close. The best private, proprietary model (Fable 5) and the best free open source one (GLM 5.2) are now separated by ~13 weeks, and the gap seems to be getting smaller with every launch.

What actually matters is the cost of running the thing, and the best way to make your model smarter is by getting it to think more before it answers. This strategy uses 5-10x more inference than models used to. Inference = more thinking time = better answers. This is irrespective of how impressive the model weights are.

So the moat has basically shifted to how effectively can your model think which means the best AI model will:

• Use inference the most intelligently

• Do it for minimal cost

Don’t believe me?

Inference is now roughly two-thirds of all AI compute and somewhere between 80 and 90% of what a model costs across its life. It passed training in data center revenue late last year.

You can see it in OpenAI’s own books. It closed Q1 on a 39% gross margin and wants to be at 52% by December, cheaper inference is the only way that gets them there. Anthropic is the same, they’re currently making 80% gross margin on every Opus 4.8 inference call, which will likely make them the first profitable AI lab in the entire world.

Jevon’s Paradox: There is Hope

The price of a token has fallen something like 1,000x in three years, and total spend on inference increased the entire way up. Every cost reduction just means an alternative use that wasn’t worth the money the week before can now be executed.

Agents are the obvious example. A single agentic task uses 5-30x the tokens of a normal chat, and in some cases while you’re asleep. Goldman sachs expects token demand to grow 24x by 2030, to 120 quadrillion a month. Does anyone seriously think we use less of this once it’s cheaper? Come on now.

NVIDIA is Sweating

Ok so if inference is where the money’s made (and how AI labs gain a lead over the others), the question becomes who gets to collect it.

Nvidia is the most valuable company in the world because its GPUs own training. CUDA is a moat nobody is crossing this decade.

But inference? Wide open, which is why the entire field is suddenly piling into it and why I think companies like Etched, Cerebras and Groq have good shots at becoming “the next NVIDIA”.

Etched claims one of its eight-chip servers does the work of 160 H100s. Groq and Cerebras are already used at scale across frontier labs and every hyperscaler is now building its own inference chip in-house, from OpenAI’s Broadcom project to Amazon’s Inferentia to Google’s TPU, purely to stop handing Nvidia its 80% margin.

Analysts think Nvidia’s share of inference falls from north of 90% today to somewhere around 20 to 30% by 2028. This sounds very dramatic but unless NVIDIA offers a meaningful chip architecture thats remarkably better than competitors in inference - they’re going to lost market share.

Bet on the Tokens

I’ve started to ignore how technically intelligent a model is in terms of weights and started paying attention to how hard the model is working to “think” and truly come up with a good answer.

Inference is the key to model thinking and the cheaper and more performant you can inject this into your models the more likely you are to win this entire race.

Bet on the tokens.

Thanks for joining us for our latest issue. Now go listen to our podcast :)

Good afternoon,

OpenAI just built its own AI chip designed entirely by ChatGPT. Then less than 24 hours later, they were told by the US government that they couldn’t release their upcoming GPT5.6 model.

Meanwhile AI memory giant Micron printed the most absurd earnings report in history (even beating NVIDIA).

And Google had the worst week a frontier lab can have, losing a Nobel laureate AND its CTO in the same week.

Anyone looking at the news headlines this week would think the government restrictions is a nightmare scenario… it is, but not for the silent winner—open source (aka Chinese model labs) :)

Let’s get into the news.

-Ejaaz

OpenAI Announced Jalapeño Before Ban

This week OpenAI unveiled its first custom AI chip Jalapeño, built with Broadcom in just 9 months. This could be the fastest chip design cycle in semiconductor history.

The chip is designed for inference and is roughly 50% cheaper per token than today’s Nvidia GPUs. The whole idea is to stop paying the Nvidia tax (currently 40-70% margin per chip) on every ChatGPT reply.

But it seemed this was too spicy of an announcement (sorry I had to) for Trump. The US has reportedly reached out to OpenAI to stagger the release of their upcoming GPT 5.6 model, citing concerns of security risk similar to Mythos.

We’ve unfortunately reached a point where models are good enough to make better chips than the #1 company in the world, but 99% of us are excluded from using the thing.

Micron’s Record-Breaking Earnings Report

Bears are cooked.

Micron did $41.5 billion in revenue last quarter, up 346% from a year ago. Sounds impressive right? It gets better: gross margins hit 85%… this is way higher than the industry average of 30%.

The stock rocketed 20% on the news. A lot of investors claiming the memory bottleneck trade was over need to figure out what their next excuse is going to be.

Next quarter Micron is guiding toward $50 billion.

The crazy part: their entire 2026 memory supply is already sold out. They also locked in 16 multi-year customer agreements that management says will eventually cover half or more of all company revenue. Oh, and they signed Anthropic this week, both a supply deal and an investment in Anthropic’s $965B funding round.

gg.

Google’s Terrible Week (and a24 Investment)

Google DeepMind lost four researchers to Anthropic and OpenAI in 7 days:

• Noam Shazeer (co-author of the Transformer paper that all of modern AI is built on) left for OpenAI.

• John Jumper (who won a 2024 Nobel Prize alongside Demis Hassabis for AlphaFold) left for Anthropic.

• Jonas Adler and Alexander Pritzel (key Gemini team members) joined Anthropic.

The exodus wiped roughly $270 billion off Alphabet’s market cap. This comes at a time where their flagship model Gemini is not actively competitive with Claude or GPT. Despite having all the capital in the world, Google has failed to catch up and the market worries it’s too late.

The irony is Google owns ~14% of Anthropic. It’s basically funding the company poaching from them.

On a more positive note, Google just wrote a $75 million check to indie studio A24 to build AI tools for filmmakers, its first-ever stake in a Hollywood firm. The deal pointedly gives Google no access to A24’s film library. A24 is now valued at $3.5B.

SpaceX’s Next Billion-Dollar Compute Deal

SpaceX just agreed to lease Nvidia GB300s to open-source lab Reflection AI for $150 million/month through 2029, a deal worth up to $6.3 billion.

If you stack that on its existing deals with Anthropic (~$1.25B/month) and Google (~$920M/month), SpaceX is now pulling in roughly $2.3 billion a month renting out compute from their Colossus data centers.

SpaceX has officially become a legitimate neocloud provider to frontier AI labs and it’s finally helping them work towards justifying a $2 trillion IPO. But it’s still a stretch and the stock price is now suffering

The rocket + Starlink business does ~$19 billion a year, and added to these these neocloud deals still puts them at 150X earnings. For context, Anthropic at $1T is still only 20-30X earnings.

SpaceX stock has crashed to IPO listing levels, and the next few months will be interesting to watch with 54% of their stock unlocking.

Meta Is So Back (9 months from now)

Against all odds, the only company standing up to the U.S. government and refusing to ban their frontier models to the public is… Meta.

Still, it would be helpful if they actually had a frontier model, right? Well it turns out they’ll be releasing a Mythos-level model in…. 9 months from now.

Which means Meta’s AI investment of $25B in the last year will produce a model that gets beaten by free, open source ones.

Nothing to see here.

Thanks for joining us for another issue. Now go listen to our podcast :)

The internet was a place filled with people.

Then the Dead Internet Theory emerged, built on the suspicion that most of what you scroll through is simply machines talking to machines, with real users quietly disappearing around 2016 as bots and algorithms filled the empty space.

The theory took hold before ChatGPT existed, describing a flood of fake content that had not been built yet.

But then we built it.

-Josh

Bots Control the Internet

The web became the largest gathering of human minds in history, and as implied, there was always a person on the other end of everything.

Cloudflare put a date on when that changed. This month the company, which observes around a fifth of all websites, reported that bots now generate well over 50% of all web requests.

The figure counts requests rather than hours, and people still own most of the time spent online, but traffic was the metric the web was built to sell, and the machines just took it.

Matthew Prince, Cloudflare’s CEO, had expected the crossover at the end of 2027 and watched it arrive eighteen months early. “Welp, that happened faster than I predicted,” he wrote.

The shift was largely caused by increasingly competent AI agents running our errands. One assistant booking a trip can load a thousand pages so you never load five.

This is relatively harmless until you give that agent a social media profile, have it pose as a human, and start trying to change people’s minds.

In 2016, researchers found nearly one in five election tweets came from automated accounts, steering a conversation millions believed was organic. I wonder what that figure is now, but hopefully the real people out there have been trained to distrust the online crowd.

Oh, How The Chess Board Turns

In 1770 a Hungarian inventor toured Europe with a chess-playing machine, the Mechanical Turk. It beat Napoleon, Benjamin Franklin, and convinced crowds a machine could think. In reality, a chess master was folded inside the cabinet, working the arm by candlelight.

In 1950 Alan Turing formalized the same puzzle: if a person with a hidden machine cannot tell it from a person without one, we call it thinking.

We passed the Turing Test a while ago and barely noticed. But now the benefit of the doubt is gone: we look at a face in our feed and wonder whether the person is secretly the machine. If someone beats us at chess, we automatically assume they used a bot.

But what most people overlook is how few real voices we ever heard in the first place.

So You Might Be Dead

The analyst Samo Burja divides the world into live players and dead players: a live player can do something genuinely new, while a dead player only runs the script. By that measure the web was always mostly dead.

In almost every online community, about 90% only watch and 1% ever create anything net new. On Wikipedia, a thousand people write two-thirds of the content seen by 4.4 billion unique monthly visitors.

The bots didn’t really kill a living internet, they just flooded one that was already mostly asleep. A large language model is the ultimate lurker, a machine that averages everything written before it, endlessly fluent and never once new.

If a machine answers your question better than the old forum did, instantly and for everyone, it is fair to ask who is harmed.

But that frames the web as a vending machine, and on those terms the machines win cleanly.

You Have a Superpower

The web was always something magic, like when you met a real new friend on an Xbox server, or a nerd on a forum helped you with a super specific problem. A perfect, preordained answer can help you, but there is no one on the other end of it.

Feed an AI model exclusively the output of other models and it falls apart, starved of the one thing it cannot generate: a person seeing something for the first time.

Genuine reactions. Biological response. Boots-on-the-ground emotional impact. Qualia.

They can’t take that away from you.

The task hasn’t changed since the Turk, or since Turing. Find the live players or become one yourself. The rest, human or machine, was always just traffic.

Thanks for joining us for our latest issue. Now go listen to our podcast :)