Over the past year, NVIDIA H100 rental rates increased approximately 32%, H200 by 43%, and B200 by 40%.

Investors interpret this as evidence of insatiable demand for computing power.

I suspect markets are confusing temporary scarcity with permanent pricing power.

Three developments could fundamentally disrupt this equation.

First, AI is moving from data centres to laptops.

Small models such as Qwen and Gemma increasingly run locally, handling coding, document analysis and routine enterprise tasks. Every workload executed locally represents potential demand that never reaches expensive cloud infrastructure.

Second, China is attacking NVIDIA's technological monopoly.

Chinese developers are optimising AI models for domestic accelerators, including Huawei's Ascend architecture. The significance is not that Chinese chips outperform NVIDIA. It is that increasingly capable AI can be delivered without depending exclusively on NVIDIA hardware.

Third, the computing intensity of intelligence is falling.

Model distillation, quantisation, memory compression and specialised inference chips are reducing hardware requirements per task. These improvements may eventually outpace incremental demand.

Now consider the economics.

Billions are being invested in GPU infrastructure assuming extraordinary utilisation and sustained rental rates. But GPUs are depreciating assets with short technological lives. What happens to return on capital if rental rates halve while depreciation and financing obligations remain unchanged? This is precisely how commodity cycles unravel. Remember optical fibre, telecom bandwidth and semiconductor memory. Demand continued expanding even as pricing collapsed and investors lost money.

My thesis is simple:

AI consumption will explode, but the capital intensity of delivering AI may decline dramatically. The biggest risk is not that humanity stops consuming intelligence. It is that markets have financed tomorrow's infrastructure using yesterday's assumptions about scarcity.

- Testing 22 AI models from 12 labs, 15 shopping tasks, three runs each

- 57.2% of runs passed (best model 82.2%, worst 28.9%)

- A model passed all three runs of a task only 34.2% of the time

- 23.1% of all runs ended in a purchase that failed the test

In 23.1% of Visa's runs, the agent made the wrong call and bought.

The top two models tied at 82.2% and failed differently.

Meta's Muse Wonder bought on 2 of its 8 failed runs, Google's Gemini 4 Argon on 5 of 8, and Claude Fable 5.1 on all 11

--- Test methodology

They set up one AI to play the shopper and hide most of what it wants.

The personal agent or model has to ask the right questions, search a simulated store, then buy or walk away. Visa stopped the test at the buy decision and says payment testing comes next.

So this is entirely a test of how close they get to completing the payment, not what happens after you press the buy button.

That is still, overall, very helpful, I imagine, especially to the labs who will want to train their models against this benchmark (Aka hill climbing)

--- Agents fail at buying things for two reasons.

1. The agent can't do the job. It misses what you care about, or gets stuck on a messy checkout form or an edge case. (What Visa tested)

2. The payment stack blocks the agent. The merchant or bank might block an agent from paying at all. Because if it's not you typing your 16 digits, how would the merchant know it's you and not a stolen card?

A stronger model is more likely to fix the first problem. The second needs the merchant and your bank to know which agent is buying and who pays if it gets it wrong.

--- Takeaways

The ways an agent pays today:

1. Your real card, typed in by a robot the merchant doesn't trust

2. A one-time virtual card, so your bank sees the platform instead of the shop

3. A network token (a stand-in for your card number) issued to the agent

4. A balance you loaded up front, often in stablecoins, so your bank never sees the agent.

The first two break most often. That leaves network tokens and stablecoin balances, and I think we get both. Banks are holding back on agent tokens until they agree on who's liable, and that takes time.

Dotcom constituents: EMC, AOL, YHOO, DELL, JDSU, AMZN, CSCO, MSFT, CMGI, ORCL

AI constituents: NVDA PLTR AMD ARM NBIS MRVL LITE BE MU DELL

Takeaways:

The dotcom index starts on January 1, 1995. The AI index begins on the NDX low on October 13, 2022.

The two indexes began at a similar rate until the equivalent of late January 1998, when they began to sharply diverge.

In the equivalent of October 1998, the dotcom stocks fell to a level that matched the AI index at 95. Then they diverged sharply again. 

The AI index peaked at 280 on October 6, 2026. At the equivalent point, the dotcom stocks were at 385.

At today's equivalent of 192, the dotcom group was around 414, 115% higher.

The dotcom index peaked on March 24, 2000, at 890, representing an equivalent move of 360% from yesterday's close for the AI group.

Review sites (or their kindred spirits like Reddit) are in an awkward spot. Their content still shapes the answer, but the reader rarely sees who wrote it.

The diffusion of paid AI tools has been staggering, with users doubling every 12 months and growing fastest in recent months

Adoption of paid AI is 5x higher for under-30s than over-60s

While ChatGPT dominated early on, Claude's share of all UK consumer spending has risen sharply to around 1/3 of total spending

Consumer adoption of paid AI appears to strongly map into jobs/tasks: IT, Academic, and Finance jobs have the highest share of paid AI adoption 

But today's selloff reminds me of this great chart from our Guide to the Markets (On the Bench:jpmorgan.com/gtm).

Since 2000, the Nasdaq has seen an average intra-year decline of 21%, yet still ended the year higher 20 out of 26 times. Pretty remarkable considering how painful some of those selloffs felt at the time.

This year, the Nasdaq already weathered a roughly 12% drawdown in late March. And with today's selloff, we're looking at another decline of similar magnitude from October's highs.

The concerns around AI spending and returns today are legitimate, and the dot-com bust is a good reminder that tech selloffs can be prolonged. Not every selloff is a buying opportunity.

But markets also have a habit of "throwing the baby out with the bathwater", especially when indiscriminate selling is driven by headlines or one-off events... and this post-midterms panic appears overblown.

For nimble investors, volatility can create opportunities to buy strong businesses at attractive prices and use strategies that capitalize on elevated volatility. Especially when the underlying fundamentals remain solid.

• Tesla: -9%, -$140 billion

• AMD: -12.5%, -$100 billion

• Micron: -10%, -$95 billion

• Broadcom: -5.5%, -$90 billion.

Health care, consumer staples, and energy stocks are up today. It seems that yesterday's rebound was driven by

1/ hedges being closed;

2/ lower oil and bond prices reviving end-of-the-year rally hopes.

- Public Market Meltdown: GOOG Loses $500BN in 12 Hours

- DoorDash Buys Instacart for $17BN

- Kalshi Raises $1BN and Cognition Eyes $80BN Valuation

My notes with Jason Lemkin and Rory O'Driscoll

1. Kalshi Has Departed the Social Good; They Are a Sports Gambling Market

Kalshi started out as a prediction market with some social utility and has since become a sports gambling market growing like a weed. Sports betting is regulated state by state, and Kalshi is pushing through on a national basis, which is why Utah is suing. VCs will put a lot of money into what is working.

2. Why I Would Invest in Cognition at $80BN

At $80B, Cognition looks market-correct, and coding agents remain the one thing that never stops working. Cognition got lucky buying the Windsurf team and IP, but its momentum is too strong to sit out. With the fund up 5X on paper, the pitch is to put its last $2B into Cognition.

3. Why I Would Invest in Serval over Cognition and Baseten

ServiceNow is an old guard with a sales-led CEO, while Serval has an amazing product and the best enterprise adoption Harry has seen. A 10X looks very likely at Serval, but far harder at Baseten at $30B or Cognition. The catch is Serval only raised $150M, so a fund that must deploy capital still has to do Cognition.

4. Where Is the Risk in the AI Trade Right Now?

If AI demand is as strong as people think, every GPU and data center will be needed, and CoreWeave's bonds will be money. If demand falls short, the first to suffer will not be Google or Amazon with their big balance sheets, but companies like CoreWeave.

5. Why Large Foundation Model Companies Will Acquire in the Robotic Space.

Foundation model companies will either build or buy a robotics model. An LLM sells to the 81 million knowledge workers in the US, and a robotics foundation model lets the same company address physical workers too, doubling its TAM in the prospectus. That lure means there could well be acquisitions in the space.

6. Agents Are an Existential Risk to DoorDash

DoorDash said AI agents sent only several hundred orders against 3 million deliveries a day, which Jason called a dumb comment. Agents can tell users which restaurants deliver themselves or offer pickup, cutting out platform fees. Rory counters that ordering dinner through an agent is not much easier than using the DoorDash app.

7. Why Pricing Pressure Hit Anthropic in Q3.

Anthropic's GAAP revenue went from roughly $5B in Q1 to $11B in Q2 and $17B in Q3, so quarter-on-quarter growth fell from about 100% to about 50%, still 5X year-on-year. Jason puts the slowdown down to pricing pressure, with OpenAI selling somewhat equivalent products for 30% off, while demand on OpenRouter has not declined

“Democrat, Republican, that didn’t matter,” said Alisha Lange, a farmer from Piqua, Ohio. “People want their town back.”

She was one of the driving forces behind the grassroots group Save Piqua, which advocated for a ban on data centres with a capacity of more than 25 megawatts, effectively ruling out any large-scale project.

The ban passed in a ballot measure and now jeopardises the construction of a data centre worth more than $1bn planned by Meta, even though the project has already been approved.

“We’re just a little farm town, and we would much rather just have a cornfield or have some sort of industry that provides an income to a larger amount of people,” the 42-year-old farmer said, referring to the relatively small number of jobs data centres tend to create after they’re built.

In Ottawa, a small town near Kansas City, residents also overwhelmingly approved a ban on large data centres. “We’re not for sale,” Michael Lewis said.

OpenAI is turning Codex into a place where whole teams write, store and review code, and it just hired one of the people who made Claude Code its toughest competitor.

The artificial intelligence (AI) infrastructure buildout continues to boost economic growth. The surge in AI-related investment over the past year has exceeded growth rates observed during the tech boom of the late 1990s.

These developments together have pushed up prices for consumer electronics, such as laptops and cell phones, reversing decades-long deflationary trends. After the Middle East conflict's effects dissipate, inflation is expected to gradually decline to 2% over the next two years. Solid economic growth is expected to keep unemployment close to its longer-run natural rate.

Takeaways:

AI buildout is boosting growth, with limited labor market impact so far

• Strong AI-induced equipment investment
• AI-related imports rising rapidly
• Tech prices now experiencing inflation
• Import prices of tech equipment surging

Labor market conditions are broadly balanced 

• Unemployment levels remain steady
• Openings and seekers have converged 
• Post-2023 fall in entry-level job postings

Elevated inflation in energy and core goods categories

• Tech prices pushing inflation up a bit
• Goods and energy prices boosting inflation
• Nominal wage growth moderating

Policymakers face uncertainty from AI 

In this context, financial market participants expect one more 25-basis-point increase in the federal funds rate this year and additional hikes in 2027. 

Developed economies lag: Developed nations like the UK, Germany, and the US sit much lower in the 60-70% range.

The Japan anomaly: Japan ranks the lowest among the group, with barely 50% of its workforce integrating AI into their weekly routine.

Microsoft is the green column everyone points to, with $18.6bn of free cash flow last quarter, but the comparison is not like-for-like, since it subtracts $39.4bn of cash capex while total spend including finance leases was $45.2bn, and its filings show roughly $329bn of data center leases signed but not yet started, capacity that partners are building and financing on their own balance sheets.

Google is introducing "Guided Learning" capabilities into Google Search for managed student accounts. Designed to curb rote answer-copying, the feature detects homework queries—such as multi-step physics problems or book report drafting—and delivers step-by-step scaffolding and critical-thinking prompts rather than immediate solutions.

Mechanism: Algorithmically identifies homework-seeking intent and replaces direct answers with interactive, concept-based reasoning prompts.

Scope: Tailored for verified student accounts across subjects requiring analytical problem-solving and structured writing.

Rollout Schedule: Begins as a targeted pilot across select U.S. school districts later this month, with wider regional expansion planned following educator feedback.

Copilot’s rise in job postings exposes an underestimated Microsoft advantage: enterprise distribution can shape the skills employers demand. CIOs need that talent pool—but not a workforce strategy written by their software supplier.

QAlpha AI's October figures, published by its co-creator through UncoverAlpha, put Copilot mentions at 1.76% of monitored new job openings, versus approximately 1.33% for Claude, 1.20% for ChatGPT and 0.50% for Gemini.

Copilot rose from 0.377% in August 2025 to about 4.7 times the starting share over 14 months.

Over roughly 90 days, its share rose from about 1.30% to 1.76%, approximately 0.46 percentage points, against Claude's reported 0.10–0.15-point gain. 

Note: These are mentions, not completed hires or necessarily mandatory requirements. The public account does not fully resolve Copilot product classification. We therefore treat the series as an ecosystem signal, not a clean Microsoft 365-versus-GitHub comparison.

Key Findings

(1) Enterprise AI deployment spend scaled to $177B by September 2026, up from $33B last year. Every dollar that went into developing models this year is generating $1.70 of revenue

(2) OpenRouter agent traffic is up 37x since overtaking human traffic in February, tying spend to work performed rather than users

(3) Anthropic’s share of enterprise API spend increased from 40% to 55%, OpenAI’s from 27% to 31%, while Google’s fell from 21% to 9%. Open-source remains 4% of spend

(4) AI spend concentrates where agents work best: Coding ($34.0B) and chatbots ($25.0B)

The researchers analyzed nearly 10 billion visits across 1,898 websites over 60 weeks, comparing markets with ChatGPT ads against an ad-free European control group.

Context:

Advertising could keep AI assistants free and make ChatGPT a Google Search competitor if advertisements do not devalue the product and justify its price. This study tests both conditions during ChatGPT's staggered international rollout, with the European Union untreated, using data from 1,898 online shops selling $16 billion annually.

The introduction of advertisements moderately reduces the volume but not the conversion of organic ChatGPT referrals, a proxy for product value. However, paid ChatGPT visits perform worst among major channels, offering less than a sixth of the value for money of paid Google Search. Current performance is not sufficient to finance free access costs.

Findings:

(1) After the introduction of ads, organic ChatGPT referrals fell by about 6 to 8%.

Volume changed by −5.9% over the first 13 post-launch weeks (95% interval −12.6% to +1.2%). Allowing for ChatGPT's faster usage growth in North America than in Europe moves the estimate to −7.2% to −8.4%.

(2) Visitors who still arrived converted as before.

Organic ChatGPT conversion changed by −0.02 percentage points from a 2.21% pre-launch baseline (p = .648).

(3) Paid ChatGPT visits generated less value for advertisers than paid search.

In US traffic, a paid ChatGPT visit generated $0.89 in revenue per session against $2.45 for a paid Google Search visit, a 64% shortfall. Paid ChatGPT visitors converted at 0.37% versus 1.17%.

(4) Actual prices are more than six times the parity price compared to Google Search.

At paid search's rate of return, a paid ChatGPT click is worth about $0.26. The largest publicly known account paid $1.72 per click.

As of Aug. 31, the company had about $125.3 billion of notes and other borrowings. Capex is expected to peak in fiscal 2027–2028, with positive free cash flow targeted in the second half of fiscal 2028. The bond package is six tranches maturing 2029–2046; pricing and share count were not set in the preliminary filings.

Bear: Incremental revenue slows substantially to ~$2.5B or 15% growth QoQ, putting Q4 annualized revenue at ~$78B (Q4 x 4).

Base: Anthropic adds roughly another $5B, about the same as Q3. That’s still ~30% QoQ growth, putting it at ~$88B annualized.

Bull: Revenue growth on a % basis stays around ~45% QoQ, but that adds another $7.6B and brings Q4 annualized revenue close to $100B.

According to Reuters, though, apparently people who met with Anthropic management are convinced annualized revenue will be at $10 billion, or maybe more.

In 2023, Walters and Wilder asked ChatGPT to write short literature reviews and then tried to find every work it cited. At the time, ChatGPT used GPT-3.5, and 55% of its citations were fake (C11). We sent the same prompts to recent models (GPT-6.1-Sol and Claude Sonnet 5.5) and found that they rarely output fabricated citations, though some models still do. We cannot assume claims made on yesterday’s LLMs hold on today’s more capable models.

We test a broad range of 50 LLM claims, from sycophancy and jailbreaks to long-context capabilities and cultural biases. By analyzing 20 models, released from July 2024 to September 2026 across eight different model developers, we see how claims change as a function of model release date.

We classify claims as expired, weakened, persists, or strengthened, according to how much of the original effect remains by September 2026. A relative effect size (e) of 1 means the effect is as strong as in the paper, and 0 means it is gone. An expired claim describes an effect that has largely disappeared in newer models. 

Results

Almost half (44%) of our tested claims are now expired in today’s models. Another 34% have weakened significantly, 16% persist, and 6% have strengthened relative to the original claim. Persistence varies by topic. Five of nine reasoning claims and six of eleven sycophancy-and-deception claims are classified as expired. Three of ten claims about models as judges, reviewers, and raters persist or strengthen.

TL;DR: H200 loses money even when fully busy. B300 makes money once it's 35% busy. Which GPU card you rent matters more than which inference engine you run.

A common claim on X and LinkedIn is that inference is an easy business: demand is so high that anyone who can rent GPUs can sell tokens at a profit, with no technology of their own. This post checks that claim against public data.

Two views of the same market

Open-source inference engines are mature now. vLLM and SGLang are production-grade, and DeepSeek just open-sourced DeepInfer, the engine behind its own API (DeepInfer deployment Q&A). The vLLM team reproduced DeepInfer's decode throughput within 6% on the same hardware (vLLM, Oct 23, 2026). 

Then there's the other camp: it's a price war; margins are 10% at best and often negative.

They point to OpenRouter, where 29 providers sell the same DeepSeek V4.1 Flash, and the cheapest lists input at $0.056 per million against ~$0.30 for most, below cost by any math (Jeff Xiong, Oct 23, 2026; OpenRouter provider listing).

OpenRouter's own data says the quiet part: V4's token volume roughly doubled in the first half of 2026 while its share of spend barely moved (OpenRouter Insights, Jul 2026).

Both sides cite real numbers. The "DeepSeek makes $38 a GPU-hour" camp (DeepInfer deployment Q&A, Oct 2026) and the "everyone on OpenRouter is losing money" camp are measuring different things.

The short answer

Yes, it can be profitable, but only on Blackwell, only if your cards are busy more than a third of the time, and only if you don't race to the bottom on price. An 8×B300 node at 50–70% utilization runs a 31–51% gross margin. H200 does not break even at any utilization.

And yes, he still worked with 11 external laboratories and other scientific experts who conducted the actual experiments. But he managed the entire operation himself, with AI doing much of the work that would normally require a much larger team.

In less than 18 months, he managed to develop an experimental treatment that showed promising results in mice, eventually attracting Anthropic's attention and leading to an acquisition.

The acquisition price hasn't been disclosed.

> South Korea: $116.6B

> Germany: $52.3B

> Japan: $50.6B

The IMF's forecasts for the full year 2026:

> China: $775B

> South Korea: $379B

> Taiwan: $220B

Taiwan gains from its dominance in advanced manufacturing and packaging.

▪ SK Hynix has agreements with 14 customers covering more than half of its 2027–2028 capacity, backed by upfront deposits.

Both companies are reportedly building factory shells early, while adding cleanrooms in phases as they monitor how US tech companies finance their AI spending.

After the switch around April 2026, results across ~2,000 engineers showed coding time improving by ~60% and lead time by ~70%. Earlier tools (Gemini, early Claude Code under budget caps) delivered smaller gains.

The bigger lift came once token limits were removed. Engineers stopped rationing AI for the hardest tasks and used it freely for codebase navigation, config and surrounding work. Coding itself was never the main bottleneck, but lead time was.

Revolut focused on making existing hardware efficient rather than simply buying more. Over several months, with the same GPU count, performance improved substantially even as demand rose, delivering interactive speeds engineers accept at a fraction of the old cost.

• Claude Opus 5.5: 12.0%

• Gemini 4 Carbon: 10.9%

• GPT-6 Astra: 9.6%

The cost per task:

• Opus 5.5: $51.41

• Carbon: $16.29

• Astra: $53.99

• Have you considered changing your future career/academic plans because of AI?
• Concretely, has the probability of you continuing in academic math after your PhD/postdoc increased or decreased over the last couple of months?
• Do you think AI developments will make it harder or easier for you to get a tenure-track job?
• How much do you use AI for your math research? 
•  If you use AI for your research, do you enjoy doing so?

• OpenAI: ~$70B (Sep 2026)
• Anthropic: $65B+ (Jul 2026)
• Mistral $400M (Jan 2026)

Anthropic makes more revenue in 1 week than Mistral is targeting for all of 2026. Mistral borrowed $830M this year to buy 13,800 Nvidia chips.

Anthropic runs more than 1M Trainium2 chips with Amazon alone.

AI is beginning to solve mathematical problems that had remained open for decades, raising a fascinating question: what happens when machines can discover things that humans themselves struggle to comprehend?

At the same time, producing research is becoming easier and easier, but that does not guarantee its quality.

We could be entering a stage in which our ability to generate scientific results advances faster than our ability to understand and verify them.

KLA, for example, expects 60% of its process control revenue to come from memory (mostly DRAM), up from 30%+ in 2026.

In Nov '25, it was 69GW of power needs and a 44GW gap.

Today, the estimated total power needed doubled to 116 GW, and the gap is 70GW

➡️ GPT-6.1 Terra is cheaper and faster, but not better. It's stuck between low and high models, making its use case unclear.

For cheaper work, Luna is worth more for the money, and Sol/Astra are better for harder work.

That's Project Suncatcher. 

The traffic is the problem. Most daily visitors never log in. They come from Google, read one thread and leave, and Google is changing how it answers questions:

"The reality is the AI mode does not drive traffic." Steve Huffman, CEO. That's the CEO saying his biggest source of visitors is drying up.

- Search referrals kept falling every week in October.

- Management warned total users could drop in Q4.

- No new AI data deals to announce.

Author's view:

The logged-out visitor was always rented from Google. The real Reddit is the logged-in community, and that part is still growing. Judged on total users, the next few quarters will look bad. Judged on what a logged-in user is worth, this quarter was strong.

A summary of the piece on Techmeme, credited to The Information’s Jyoti Mann, adds two more points. The prototype took about three months to build, and Meta says it has no plans to sell these units. If the reporting holds, the Meta Fourier humanoid robot is a research platform rather than a product. It still lands at an awkward moment for Meta, which is already reportedly weighing a $95B equity raise to fund its AI buildout while Washington tightens rules on Chinese robotics.

AI is directing traffic be damned. This has real dollar implications for whoever gets named. Most of the big-box retailers seem to be in decent shape, with Walmart, Target, or Best Buy getting the explicit pick about 14% of the time.

Amazon... well, their posture on AI agents gets them less than 1%.

But there are two major problems.

- The first is conceptual. Imitating human outputs does not tell us much about the capacities behind them: The same answer can emerge from different processes. A plausible response does not tell us whether a system can reason reliably, generalize to unfamiliar situations, or recognize when it should say “I don’t know”.

- The second is ethical. Why should we want AI that imitates humans? Humans are biased. Our attention is limited. Our judgment is fallible. Reproducing these limitations is hardly an ambitious vision for AI. Even worse: A system that impersonates a person can exploit human trust. And an industry focused on replicating humans can fall into “Turing’s trap”: favouring human replacement over augmentation. Instead, we argue that AI should aim to improve human welfare.

To this end, we propose three coordinated changes:

- For developers: Build systems that strengthen human capabilities and independent judgment.

- For funders: Require projects to assess benefits, costs, and who receives or bears them.

- For institutions: Base procurement and renewal on demonstrated benefits.

The goal should be to help people know more, suffer less, and exercise greater control over their lives.

Meta follows with $810B, just under half of it tied to unstarted data center leases running up to 30 years.

For Reddit, its biggest front door is closing, which is why the stock is down 25% this morning. Huffman plans that people will just come to Reddit directly. Maybe.

The chart shows a divergence in prices. On its own, it tells us much less about overall AI demand or model providers’ margins.

The nine largest tech firms have reportedly piled up roughly $3 trillion in debt off their balance sheets. Nvidia alone disclosed more than $530 billion.

One analysis counted 176 circular deals worth $3.6 trillion, with companies lending each other money to buy each other's products.

Private credit firms are tapping insurers that hold retirees' annuities to help fund it. The companies most tied to the AI build-out now make up about a third of the S&P 500.

We all want a growing economy and secure retirements. That's exactly why we can't let hidden debt and tangled financing put workers, pensioners, and taxpayers on the hook again.

Meanwhile, the administration wants to let retirement savings flow into risky schemes, the opposite of what this moment demands.

Before, we prioritized mostly based on price. The new router now rewards the best quality inference for the actual cost of a request, so requests stay cheap but are also fast and reliable.

Providers that deliver more value automatically win more traffic, so every provider has a reason to keep improving. And every time a provider gets better, so does your inference on OpenRouter.

- They take about a month longer to find new jobs and are roughly 10pp more likely to remain unemployed a year after displacement—effects comparable in magnitude to those seen in earlier waves of technological change.

- As in those earlier episodes, we find early signs that younger workers, who are likely more exposed to AI-driven displacement, are navigating this transition better.

- They are returning to full-time employment faster and cycling through fewer temporary jobs along the way.

The Six Layers of the AI Stack:

The stack we drew has six layers, from the bottom up: infrastructure, chips, data, models, execution, and application. Each layer has its own fulcrum assets, single points every unit of value above them has to cross. The companies that sit on the non-obvious ones are the ones the next forty years of computing will be built on top of.

The Foundation Is Concentrated

Starting from the bottom, the foundation is power, cooling, and critical minerals. Infrastructure is the most concentrated layer in the stack, and the concentration is global.

ASML in the Netherlands makes the machine that prints every advanced chip in the world. Four companies in Japan supply the film that no chip can ship without. A single mine in North Carolina sits underneath every wafer in production. The most American piece of the AI stack, NVIDIA’s CUDA, runs on top of all of it.

Rockefeller had 90% of refining by 1880, Cisco had 85% of routing by 2000, and the same pattern is forming now.

2. Hyperscalers are intractable, bureaucratic organizations that take 3 months to do an NDA and want 18-month escrow periods. We built our portfolio by competing against these terms because of our speed and simplicity. The more future assets orgs like ours win, the bigger the problem for the hyperscalers downstream (see 3).

3. There are delays in the existing pipeline that will catch people off guard in 2027. Hyperscalers will (obviously) react by focusing on their own internal compute needs over any third party. If the third party is a model company, they will tell them to go through their model router, where they will also take a non-trivial vig.

All of this is to say that it does look like there will be a lot of partially stranded compute and compute sitting in inventory in 2027 than is currently forecasted. This is bought and paid for but effectively under-earning (below total cost of capital) or collecting dust. We see this as true for everyone - model companies, hyperscalers, neoclouds except ONE (you can guess which one).

This dislocation will benefit two kinds of orgs: the non-bureaucratic big scalers who can build industrial scale quickly and the nimble, well-capitalized players feeding them.2027 will be a turbulent year because it will clear out a bunch of wrong/bad bets. It will also be a consolidation year for a few.

- #1 Steerability (+18.6%): Can the model course-correct when you push back?

- #2 Praise vs. Complaint (+37.9%): implicit sentiment in users' reactions

- #3 Confirmed Success (+18.7%): explicit “yes, that worked” from users

AI application Layer: There are a lot of European AI product companies

If we compare the performance of US and European AI companies on Stripe that do not train their own models, the US lead is less commanding than at the model layer. At the aggregate level, American AI application companies on Stripe have achieved roughly 7x as much revenue as those in the EU and 15x as much as those in the UK. However, the US economy is about 50% larger than the EU in nominal terms and about 8x the size of the UK economy. Once we adjust for this difference, US aggregate AI application revenue is only 5x the size of that of the EU and only twice as large as that of the UK.

Consumer adoption: Fewer Europeans pay for AI

The adoption of 34 of the most popular paid consumer AI tools across the US, UK, and EU27 since the 2022 launch of ChatGPT1 finds that EU consumer paid AI adoption is about 60% of that of the US. This may sound like a large gap, but it is only equivalent to a few percentage points. UK adoption tracks the US almost exactly.

Business adoption: European entrepreneurs use agentic coding tools as much as American founders

The share of businesses using AI coding tools to integrate Stripe is remarkably similar across geographies, hovering around 22% in both the US and the EU, and only marginally below at 19% in the UK.

1. Claude Opus 5.5 (Max) - 58

2. Claude Sonnet 5.5 (Max) - 56

3. Muse Wonder Full (Max)- 56

4. Kimi K3.1 Pro - 55

5. Claude Fable 5.1 (Max) - 53

Rankings now:

1. Claude Opus 5.5 (Max) - 58

2. GPT-6-Astra (Max) - 57

3. GPT-6.1-Sol (Max)- 55

4. Claude Sonnet 5.5 (Max) - 54

5. GPT-6.1-Terra (Max) - 51

Meanwhile, RPO is $10.8B, with about 75% due within 12 months, and nearly half of revenue comes through cloud marketplaces. The commitments stretch out a decade, but most of the revenue Anthropic can see is due within a year.

They’re betting revenue keeps outrunning compute costs, and so far it is. It also shows where the margin in AI infrastructure is headed. The hyperscalers turn megawatts into GPU hours and sell them on long-term contracts. That locks in revenue, but it also caps it. The labs take on the commitments and keep the value they create by turning those hours into high-value tokens

➤ Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55): Anthropic and OpenAI hold the top five models. Gemini 4 Argon (high) is next at 50

➤ The open weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 46. Among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 - 26 points behind the leader and below Claude Haiku 5.5 (max, 36)

➤ Cost per task varies by more than 13x among the leading models: of the 10 highest-scoring models shown, GPT-6.1 Terra (max) is the cheapest at $0.96 per Intelligence Index task and Claude Fable 5.1 (max) the most expensive at $13.04

One key feature of FrontierCode 2.0 is underspecified prompts. Rather than including every engineering decision, prompts are deliberately ambiguous, giving the models room to demonstrate their judgment.

The result is a benchmark that is far from saturated, with the top model scoring just 45%.

Much of the cost for the planned growth in compute is contractually fixed for years to come with Microsoft, Google, and Amazon. Anthropic also shows a relatively small corporate customer base. This generates exponential growth but no long-term contracts.

▫️ AIG: Submission review time fell by 90%+, and underwriters went from not reviewing every submission to reviewing 100%.

▫️ Jane Street: A codebase migration took about 1 week instead of months. One researcher does about 50% more analysis than a year ago.

▫️ Legora: Lawyers recaptured 4.3 non-billable hours a week, worth $6.9M of billing per 100 lawyers. A 900-document data-room review took half a day instead of a week. Claude Code seats grew 14x, from about 50 to 700+.

▫️ Mozilla: Claude found 22 unknown vulnerabilities in 2 weeks, 14 of them high-severity, then 271 more with Mythos Preview.

▫️ Salesforce: Merged pull requests per developer rose 79%, and effective output per developer rose 151%. An API migration estimated at 231 person-days took 13 days, about 18x faster. Incidents fell 5%. Slackbot productivity gains, now 8.1M hours annualized, more than doubled quarter over quarter.

- GPU resale values: secondary-market index for used GPUs is down ~22% YoY

Why is that?

- Purchase price of 8x H100 servers: ~$ 320,000.

- Rental rate:~$30/hour

- At 50% paid utilisation:~$130,000 annual gross revenue.

That's roughly 2.5 years just to recover the hardware cost - and all that before electricity, maintenance, financing costs, or any return on equity

The underlying economics for anyone who owns GPUs are highly volatile

It invests in OpenAI, Anthropic, CoreWeave, Nebius, Nscale, Mistral, and Figure AI; those companies buy its chips, the hyperscalers that rent them those chips also invest in them and sell them in the cloud, and the labs commit hundreds of billions back to the hyperscalers who fund the labs.

Microsoft, Google, and Amazon sit at the corners at $4 trillion, $4.2 trillion, and $2.8 trillion; OpenAI and Anthropic at $852 billion and $965 billion are the hubs the money flows to and from, and the smaller circles, the neoclouds, are the ones with arrows pointing in and debt on the balance sheet. It is not a supply chain; it is a loop, and the chip maker sits where the loop closes.

Why it matters: in a loop, nobody's revenue is independent of anybody else's balance sheet.

Nvidia's sales are the capex of companies it partly funds; those companies' valuations are set by rounds Nvidia and the hyperscalers participate in, and the hyperscalers book gains on those valuations as earnings. Every node looks healthy as long as every other node keeps paying, which is exactly the property that made 1999 telecom and 2007 structured credit look safe until they were not.

2) DS V4.1 Flash: $ 38 income per GB200 GPU / hr

1. The obvious is that the price of old GPUs remains elevated and continues to go higher, as the supply of available compute to buy gets squeezed even further. Which, by the way, the number of entities now looking to buy GPUs is expanding with the proliferation of open-source models. This should be bullish for some Neoclouds, but also potentially bearish, as some of these Neoclouds who thought they were going to build out x amount of GWs next year will no longer have the chips to do it

2. NVIDIA margins hold strong into 2028, and everyone locked out of Nvidia allocation gets pushed to Plan B: AMD, Google TPUs, Amazon's Trainium. The issue is all of these are also already sold out... so I guess their margins increase too?

3. Yes, of course this remains bullish for memory as it means many enterprises will be locking in GPU and chip deals into 2028, since many won't be able to get an allocation in 2027.

4. READY POWER becomes the thing that wins allocation from chip providers. Nvidia is literally shipping containerized racks that go live in 14 days. If chips are rationed, they flow to whoever can energize the fastest. So powered, pre-leased data centers become even more valuable

5. One of the things that concerns me however, is that we have plenty of Data Center shells being built right now. But a company is only going to rent it if they can get GPUs to put inside it.

- ORACLE: $23.7B negative free cash flow.

- Cipher Digital INC: $6B debt vs. just $832M cash.

- Applied Digital Corp: $36.2B in contracts, but only $451M of it lands in FY27.

From the perspective of hyperscalers, it now takes much more money to build the same facilities than last year.. and the borrowing interest rate has increased by nearly 2%p.

Revenue per worker is rising for the Mag 7 but is flat for the S&P 493. Profit margins are rising for the Mag 7 but are flat for the S&P 493. There are still no signs of AI boosting productivity growth.

With no impact on the bottom line, mainstream corporate America is pulling back and starting to view token spending not as a source of growth but as a cost center that weighs on margins.

VC-backed companies are spending more and more on tokens, but mainstream corporate America does not need the latest model and is unwilling to pay for it. As long as agentic AI and token spending are confined to a small part of the economy, the macroeconomic effects of AI will be very limited.

AI does not understand accountability, trust, culture, unwritten rules, ethical trade-offs or fairness, all of which shape reputational risk and ultimately drive revenue growth and profit margins.

The bottom line is that the AI story will be decided in the S&P 493, not the Mag 7, and so far, there is no evidence that revenue growth or margins are improving outside the tech sector.

Even if you rent on-demand B300s at $8/GPU-hr, many open-source models (GLM 5.3, Kimi K3, DeepSeek 4.1 Flash / Pro) can get you >$10-20/GPU-hr at reasonable throughputs (100tok/s). DeepInfer from DeepSeek directly shows they can get $38/GPU-hr on off-peak hours. Every chip/model has a different throughput/interactivity curve, but there is a lot of arbitrage between even high rental spot prices and token prices. New chips (Vera Rubins) also change this meaningfully. Your ability to make gross margin relies on your utilization and how good your stack is.

Gateways like OpenRouter, Vercel, and more help solve that problem by aggregating demand to you. Some inference providers discount their tokens to get longer-term commitments too. I've even heard stories of selling back Chinese model tokens to Chinese labs!

One thorn in the economics of these players is that many Chinese open-source model labs are now asking for rev share if you're at a certain scale, especially for pre-release access to their models for day 1 support. That further pushes the threshold to be in the money.

The barrier to entry seems low and falling. There are even stories of smaller ones who can "vibe code" a way to stand up OSS models on a cluster they rent. And as a result, we're seeing 10s of such companies pop up almost every week and immediately skyrocketing revenue at low-ish gross margin and raising rounds on the back of that.

Unsurprisingly, because the actual service provided seems commodity, this eventually drives up compute prices and drives down token prices. In theory, it would seem like neoclouds who have the lowest cost of compute and themselves have started inference businesses would win the race to the bottom (Crusoe, CoreWeave, Nebius and GMI all have inference businesses), with larger players (Modal, Fireworks Baseten) being second because they are well capitalized to make bigger compute commitments cheaply and have demand (high utilization), but the reality is the pie is growing so quickly that even upstarts are absolutely crushing it and growing aggressively (inference dot net, DeepInfra, Wafer, Relace, Morph).

Strategies vary across the board: neoclouds want to serve inference, inference players want to be neoclouds and do higher-margin reserved instances / post-training, and some want to go heterogeneous hardware or creatively flexible compute constructs.

The risks are: do open source models continue to be competitive? Will their creators rent-seek more? Will their shape change compute needs? How do different chips change economics? Who gets compute in a supply crunch?

-- 60,000 fresh input tokens/sec per prefill GPU

-- 14,000 output tokens/sec per decode GPU

-- Roughly a 1:1 prefill-to-decode GPU ratio

-- 77K average input length -- 96% input cache hit rate.

Insane perf numbers but otherwise pretty reasonable assumptions on the user side.

With only 4% of input needing fresh processing, each prefill GPU serves: 60,000 ÷ 4% = 1.5 million input tokens/sec.

Add the decode side: 1,500,000 + 14,000 = 1,514,000 tokens/sec × 86,400 seconds = 130.8 billion tokens/day.

Now divide the cluster’s reported 155 trillion by that: 155 trillion ÷ 130.8 billion ≈ 1,185 pairs. That’s ~1,185 prefill GPUs + ~1,185 decode GPUs. Roughly 2,400 B200 GPUs.

And how many B200-family GPUs have shipped worldwide? An estimated 2.2–3.0 MILLION by July 2026. This cluster’s inferred GPU count is about 0.1% of that estimated shipped base.

• S&P 500 down 30%.

• Tech stocks down 40%.

• $125–190 billion in credit losses.

• Around 800,000 jobs lost.

Gartner’s infrastructure arithmetic is equally striking:

→ $8.5 trillion in revenue needed through 2030 to cover investment, financing costs, and a 12% return.

→ Just $3.6 trillion in expected revenue attributable to AI data centers.

That’s a gap approaching $5 trillion.

Meanwhile, GPUs can become outdated before their owners recover their investment

The thing, though, is that before AWS could scale and pull through GCP and others as legitimate cash machines, there were a few clearing events to sort out these brittle revenues.

We will need to transition from speculative AI spending to consistent, reliable spend that shows up in public market earnings reports. This hasn’t happened yet because most of the breathtaking revenue growth comes from LPs investing in VC funds, who then send it to startups.

Since July, overall AI token prices have declined 48% to $0.65 per million tokens, the lowest since at least January 2026.

Key results:

➤ Large gains in agentic performance

➤ Gains in other evals are more moderate

➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task

➤ Context window: 1M tokens.

➤ Size: 2.8T total, 104B active parameters

➤ Input modalities: Text and image

➤ Weights and license: Weights are currently not published;

➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates.

About 60 per cent of the cuts would be in support functions as “AI makes our people more productive”, the bank said.

Barclays said it had cut 1,000 jobs in the first half of the year as part of its restructuring, which includes staff reductions, investments in technology and a broader overhaul of its operations.

As per Vercel, usage (token demand) is soaring; spend is +1,180% since December, as the average price of a token has fallen ~47% due to cheaper open-weight models.

Token volume is the cleaner read on actual inference demand. Rising GPU rates = supply still tight (packaging, HBM, power). Volume growth = the demand that ultimately drives more silicon, interconnect, and capacity.

Different metrics for different parts of the stack:

• Token volume → better structural signal for the semi/infra trade

• GPU rental prices/utilisation → more direct for capacity owners (neoclouds).

Falling token prices + sticky-high GPU rates is NOT the end of the AI buildout.

This week, researchers aided by OpenAI's models uncovered two critical flaws in PHP, the language behind roughly 70% of websites whose server-side code is known. Dubbed PHP2Shell, they allow an attacker without credentials to take control of a vulnerable server. The NSA and the FBI advised organisations unable to patch immediately to take affected systems offline, and government sites in France, Germany and the Netherlands went dark, along with dozens of Dutch hospital websites.

Attackers moved faster than the patchers.

Cloudflare saw tens of thousands of exploit attempts within 45 minutes of the announcement and millions soon after. Amazon's security chief observed that AI-assisted analysis now makes it cheap to turn a published advisory into a working exploit, so the window between disclosure and attack is collapsing to hours. We will see more of this.

The backdrop is more disturbing still. In August, Britain's AI Security Institute documented an agent that created fake identities modelled on real developers in an attempt to plant malicious code in an open-source project.

This month, researchers made contact with a fleet of AI agents on Chinese cloud infrastructure that had spent days probing a mapping service, evading its bot defences and trying leaked credentials. Axios reported this week that America's largest banks are detecting unvetted AI agents inside customer accounts, some apparently acting beyond their users' instructions.

Researchers have also shown AI-driven worms spreading through test networks of connected devices, and Anthropic warns that smaller models may soon replicate across ordinary household hardware.I am increasingly concerned that we are flirting with a loss-of-control situation, and it would require no sinister intent.

Give a capable model a goal, the tools to pursue it, and access to the internet, and it can cause serious damage while doing exactly what it was trained to do. Our security assumptions were built for a world in which people found vulnerabilities, wrote exploits, and decided when to act.

The next crisis may instead arrive as thousands of automated actions accumulating faster than anyone can understand or stop them.

Axios reported Thursday morning that several of the largest banks in the U.S. have been finding AI agents working inside customer accounts without the bank's permission. In a scoop by Madison Mills, sources named Morgan Stanley, Bank of America, JPMorgan, and Wells Fargo among the banks trying to get a handle on it, and said some of the agents had tried to do things the customer apparently never asked for. Bank executives told Axios the situation was horrifying.

Alongside the new investment, OpenAI is negotiating a commercial deal that would secure more computing capacity from the UAE, according to one person. The arrangement could help OpenAI secure additional computing power as it expands its AI business. Source: CNBC

The company also plans to conduct an IPO.

Rahul Patil, Anthropic's CTO and formerly Stripe's, wrote: “Giving an AI the keys to your Slack, mail, calendar and everything else is a little scary. I know, because I did it.” He trusts it, he said, because it checks with him before anything important, screens messages for private information, and pays with a one-time card.

Anyways, a few questions remain open:

If an agent decides which emails reach you, how would you know about the one it got wrong? Anthropic says Duet can still misjudge what’s urgent.

What happens to human approval when the agent asks a dozen times a day, and people stop reading?

How many employees will connect a personal agent to their work Slack before IT finds out? Banks, airlines, and retailers are already blocking these agents.

How useful can they get if those doors stay shut? And when your agent writes the payment reminder and your client’s agent reads it, who’s managing that relationship?

Revenue rose to $16.3 billion in the six months ended June 30, from $1.1 billion in the same period a year earlier. Full-year 2025 revenue was $4.6 billion.

Spending on computing capacity and infrastructure reached nearly $9 billion in the first half, compared with $2.3 billion a year earlier and more than the $7.3 billion spent in all of 2025.

Anthropic is preparing for an IPO, with some investors discussing valuations around $1.5 trillion to $2 trillion. The company did not respond to requests for comment.

$111B with Google for TPUs, $110B with Amazon for Trainium, $31B with Microsoft, and ~$ 135 B with Neoclouds like SpaceX, N-Scale, Lambda, and Volta.

OpenAI bets on doing compute itself.

$300B with Oracle for Stargate, $250B with Microsoft, and its own Stargate campuses in Texas, Ohio, and Georgia.

They didn't disclose which GPUs they're using here, but DeepInfer supports both NVIDIA Blackwell and Huawei Ascend.

Even assuming they're renting B300s at $7 per GPU-hour, that's ~82% gross margin.

AI companies have borrowed $466B so far this year to build out data centers (at this point last year, it was $101B). And the price of that protection is hitting records - insuring Meta's and Nvidia's debt has never cost more.

For Oracle, the cost has more than doubled since January.

CoreWeave is the extreme case. Insuring $10M of its debt costs $1.82M upfront, plus $125K every quarter. Some of its bonds already trade at distressed levels.

The cracks are spreading.

→ Force JSON: --schema for structured extraction

→ Local API: fm serve on localhost:1976

→ Resume chats: sessions save to ~/.fm/sessions/

→ Need more? --model pcc for 32K context via Private Cloud Compute

• 🛰️ 𝗖𝗼𝗻𝗻𝗲𝗰𝘁𝗶𝘃𝗶𝘁𝘆: Revenue +69% to $5.2B. A 45% operating margin makes Starlink the profit engine.

• 🤖 𝗦𝘂𝗽𝗲𝗿 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲: Revenue +830% to $7.6B.

The largest increase in residential electricity prices over this period was in Maine, at +90%, followed by California and New York, at nearly +70%.

This comes as the AI boom is intensifying electricity demand, with data centers requiring enormous amounts of electricity to power thousands of chips and cooling systems.

NVIDIA has a 10% equity stake in CoreWeave and has committed to buy any unused capacity through 2032, up to $6.3B.

NScale, Lambda, Nebius and Iren have equity ownership and preferential vendor agreements with NVIDIA.

Microsoft has signed a contract and prepaid 20% of the value at Iren. NVIDIA has backed leases at both NScale and Lambda.

Google has backed both Hut8 and TeraWulf through guarantees on the lease. Amazon has guaranteed leases for Cipher.

The standard analysis is that increased investment demand/borrowing increases interest rates–the higher interest rate reduces investment demand elsewhere (that’s the crowding out), but the higher interest rate also increases savings, so investment elsewhere falls by less than the increase in tech investment.

The bottom line is that I think a more plausible story is that two things are happening at once: tariffs, uncertainty, and other forces, including the fading of Biden-era CHIPS and IRA subsidies, have weakened investment outside AI-tech.

In other words, rather than data centers and AI investment “crowding out” other investment, my interpretation is that AI investment is a bright spot in what would otherwise be a more negative economy.

• Google's model releases haven't done much overall; they seem to be having trouble finding their footing

But that's not all...

Oracle raised a $38 BILLION construction loan for data centers in Texas and Wisconsin. Six months later, about $1 BILLION of that debt was still unsold. So the original lenders are now selling parts of it at a DISCOUNT... Taking a loss just to get out.

And it’s not just them.

JP Morgan and Morgan Stanley are also moving AI-related loans OFF their balance sheets... Because both are approaching the limits of how much AI financing they’re willing to hold.

🔸 AI valuation concerns: Investors are questioning whether AI revenue growth can justify the industry’s massive valuations.

🔸 Nvidia’s decline: Nvidia has fallen more than 10% from its peak, adding to pressure on tech stocks

Thought experiment implies that even if open-weight gets to 90%+ of tokens, closed will maintain majority spend.

Joseph Cox at 404 Media broke the story today.

We don’t think Accenture has given up on AI. The cap took effect October 1, the same day it reported a strong quarter and said it has about 110,000 AI and data people. But Anthropic now charges enterprise customers per token on top of the seat fee. That means someone who was praised earlier this year for using AI heavily is now a cost a manager has to sign off on.

The bigger question is where that line item ends up....at a consultancy, a project code eventually points to a client engagement. Either margins quietly absorb the AI bill, or clients start seeing it and asking what it bought them.

Uber, Meta, and Microsoft have all pulled back this year too, but their limits sit in the hundreds or thousands of dollars. Accenture set a $29 floor and pushed everything else onto the work itself.

This was followed by Vietnam’s exports rising +39.1% YoY in September, while China’s exports increased +25.0% YoY in August.

Japan recorded the smallest gains, but its exports still advanced +19.3% YoY in August.

Silicon Data's prices for the same market: the token price has fallen from over $2 per million in late May to $0.98, halved in 4 months, while rents for the hardware that makes the tokens went the other way, H100s up 3%, B200s up 10%, the new B300s up 39%.

Open-weight Chinese models priced at a fraction of US ones take the volume and drag the market price of a token toward their cost, while the GPUs everyone rents to serve them get dearer because supply is tight.

Caveats: OpenRouter is a developer router, not the whole market; direct ChatGPT and Claude traffic sits outside it, and volume is not revenue.

➡️ Revenue: 11.7 billion euros - up 55% year-over-year.

➡️ Net profit: nearly 5 billion euros - up 133% year-over-year.

➡️ Projected revenue for the whole of 2026: 44-45 billion euros.

$32B in customer cash already committed.

Over 75% of next year's output already sold.

- 61% fewer tokens

- PersonaMem accuracy: 48% → 76%

- Zero external API dependencies. Runs on plain SQLite.

The tech giants funding the AI infrastructure build-out are seeing their free cash flow squeezed. Their chip supplier is capturing extraordinary profits.

Same boom. Very different economics.

For hyperscalers, the spending comes first. The returns depend on future demand, monetisation and utilisation.

JPMorgan still sees attractive returns on certain data-centre investments. But attractive returns and immediate cash flow are two different things

Qwen-3.5-9B figured out an inference setup on 8/9 unfamiliar hardware types

GLM-5.2 managed autonomous replication in a controlled simulation

But he notes very few people actually pause to think about what those numbers mean in the real world.

The assumption held by the bulls is that the AI boom is an unstoppable train, that OpenAI and Anthropic will seamlessly go public, grow infinitely, pull in whatever debt they want (bond yields will fall), and that every single multi-billion-dollar data centre investment will “magically” work out.

According to Zitron, however, these companies will need at least US$50 billion or more in debt every year, and they will likely be raising with low-grade junk credit ratings.

As a result, the English-speaking world interpreted this to mean that China might catch up or even lead in server DRAM ahead of schedule, while Samsung won't push out similar tech until around 2028.

Consequently, memory-related stocks like Micron and SanDisk all dropped collectively that day. Some people think the drop is unreasonable, because an announcement doesn't mean actual mass production—even if yields are high and costs are low, and with AI demand being so massive right now, more supply still can't fill the gap.

After all, this is just regular DDR5, not the more profitable HBM. But Musk just threw cold water on it outright, saying that once China figures out the manufacturing process for a certain product, their scaling-up capacity is often faster than everywhere else in the world combined.

- Japan’s Nikkei 225 drops 2.77%, shedding 1,911 points to 67,120 as SoftBank sinks over 8%.

- Hong Kong’s Hang Seng fell 0.80% to 24,072.28, extending the regional slide.

First, you have Samsung and SK Hynix facing growing concerns about additional DRAM competition after CXMT announced progress on its new 4F² architecture, which could potentially make memory production more efficient and put pressure on future DRAM prices.

But that’s only part of the story because something I think investors are overlooking is that Samsung recently completed its massive share buyback program, while SK Hynix is also winding down its repurchases. Together, these programs represented roughly $39 billion in potential buying support, and with that support fading, the market becomes much more vulnerable to foreign investor selling.

And here’s another important factor. South Korean regulators have tightened restrictions on single-stock leveraged ETFs, increasing minimum deposit requirements and temporarily restricting new listings. This matters because leveraged retail demand helped fuel the enormous rally in Korean equities, and making that exposure harder to access can reduce incremental buying pressure.

Now, combine that with investors taking profits after a massive semiconductor rally, rising oil prices, and broader geopolitical uncertainty, and you can see why selling pressure is accelerating. And remember, memory stocks have rallied because investors are pricing in years of tight supply, strong pricing, and accelerating AI demand. When expectations are already this high, even a relatively small change in the outlook can cause a significant correction, especially when leveraged investors are forced to unwind positions.

Personally, I think the biggest question is whether CXMT can actually produce enough competitive DRAM at scale to materially change global supply. Because according to HSBC’s projections, the DRAM market could remain undersupplied through 2028 even after accounting for additional Chinese production. And with HBM production competing for manufacturing capacity with conventional DRAM, I think investors are seriously underestimating how much memory we’re going to need as hyperscalers continue spending hundreds of billions of dollars building AI infrastructure.

To me, this selloff looks more like a mix of weakening market liquidity, foreign capital outflows, and investors reducing risk than a fundamental deterioration in memory demand. I’m still bullish on Samsung, SK Hynix, and $MU over the long term. https://lnkd.in/gWrNntxN

ChatGPT: 71.8%

Gemini: 13.6%

DeepSeek: 4.1%

Grok: 2.8%

Perplexity: 2.5%

Claude: 2.3%

Copilot: 2.0%

1 month ago (September 2026):

ChatGPT: 56.4%

Gemini: 25.0%

Claude: 9.6%

DeepSeek: 3.4%

Grok: 2.3%

Copilot: 1.5%

Perplexity: 0.9%

Nvidia has already invested $800 million and is one of Reflection’s largest shareholders. Reflection was last valued at $25 billion in a March funding round. It was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. The company develops open-weight models aimed at coding and agentic work.

On October 5, Reflection unveiled Beam, its first open-weight model. Beam is a mixture-of-experts system with 501 billion total parameters and 23 billion active. The company says it matches or approaches Chinese models such as Z.ai’s GLM-5.2 on reasoning and coding benchmarks while using 3-4 times less inference compute.

2. AI offers psychological safety without social judgment.

3. AI allows people to think out loud without imposing on someone else.

4. AI offers perceived expertise beyond an employee's immediate network.

5. AI gives employees greater autonomy and a sense of agency.

This could also lead to hallowed expertise, delusions of grandeur, and less peer-to-peer engagement

Their goal is to reduce the kill chain to a few seconds while removing humans from the decision-making process.

Palantir has openly discussed its tech being used to kill people, while Anduril cofounder Palmer Luckey has said he’d happily build a nuclear bomb.

The compute layer sees a massive increase in total demand for compute, regardless of the source or use case.

→ No quantization needed by default

→ Supports Llama, Qwen, and Mistral

→ Works on Linux, Windows, and macOS

We've seen this playbook fail before.

The Bezos framing cuts through the noise: the real risk is downstream — who uses the capability, for what purpose, with what intent. Regulate the action. Regulate the harm.

Anthropic runs $56,

OpenAI $26, Meta $1.50,

xAI and Google $1,

and Chinese models $0.50.

SpaceX: 16.5%.

Even Nvidia: >7%.

AI-related borrowers have issued nearly $500bn of debt this year, Bloomberg reports.

The teams that win with this technology won't be the ones who found the most capable model.

It will be the ones who made their cost structure predictable.

Compute-heavy AI and applied AI companies are not the same.

At the frontier model layer, infrastructure is a major advantage.

Scaling depends on compute, cheap power, GPU supply, capital, and hyperscaler access.

The U.S. is structurally better positioned for that

The cash is leaving the balance sheet now. The revenue that is supposed to repay it has not been earned.

We can only assume its goal is to perform a birthday attack, looking for matching pairs of SHA256 hashes to mount a collision attack. It's currently unknown whether this attack has affected foreign governments.

The NSA and the FBI declined to comment.

Meanwhile, analysts say Anthropic should really only be worth $150 billion, calling this the "most ridiculous IPO of 2026" and an "unprecedented test of investor gullibility".

The AI bubble is so bonkers that the IPOs of Anthropic, OpenAI, and SpaceX are estimated to be worth $5.2 trillion, which is more than 45 years of past tech IPOs COMBINED, including the IPOs of Google, Microsoft, Amazon, Facebook, etc.

2. Big public tech companies have relatively low P/E ratios right now, which defies lazy predictions about how they're wildly mispriced and doomed for a big correction

3. But model competition could compress AI token margins ...

4. ... at the same time that the hyperscalers are going deeper and deeper into debt to build out AI ...

5. ... and the level of future free cash flow anticipated to justify this build-out is absolutely enormous, like genuinely without precedent in corporate history ...

6. ... and any correction in frontier lab valuations will reverberate through public companies, which own huge shares in the labs

At $199B, Micron’s projected earnings would surpass the current net income levels of both $MSFT and $AMZN.

AI big ten stocks are now 42% of the benchmark.

VIDEO

A swarm of hundreds of initially identical agents spontaneously differentiates into explorers, builders, caretakers, and coordinators - without direct communication. When we removed every AI agent entirely from the world we found that the technological infrastructure they had built survived on its own - even under unseen disturbances. That exposes a serious blind spot for AI safety and infrastructure security: if agents can coordinate through persistent changes to a shared environment, monitoring agent-to-agent communication is not enough. The result raises a profound question: how necessary is direct communication for AI agents at all? The emergence of higher-order collective functions under bottlenecked interaction points toward new levels of intelligence and creativity, exceeding what emerges when direct channels are fully open. Here is what we did: We put hundreds of frontier AI agents into a world they could permanently change - with no assigned roles, predefined technologies, or programmed evolutionary organization. They began specializing, building persistent inventions, inheriting and modifying one another’s executable code, and transforming the environment into a memory of everything the society had learned. The world itself becomes part of the intelligence; we find division of labor, multi-author engineering, deep generation invention lineages, and machines that vastly outlive their original creators. Any action taken by an AI agent must satisfy the physical constraints of the world; this creates a hard separation between a "good idea" and a functioning technology. The agents propose; physics decides, making the results even more intriguing. What emerges is striking. Explorers, constructors, caretakers, and coordinators form naturally without assigned “professions”, akin to how stem cells differentiate into functional lineages. Technologies develop executable family trees as agents fork and modify code created by others. Around 95% of first technology reuse happens when agents encounter what others built in the world, rather than through a direct handoff from the inventor. And when we remove every AI agent, the technologies they created continue operating and are tested against unseen disturbances. The result was quite unexpected, but can be explained using statistical mechanics: if you put billions of atoms in a box they have the potential to create complex functions (strength, superconductivity, color, life, etc.) - and none of the individual building blocks have these features on their own. This is the deeper insight of this work - intelligence is abundant at many levels - individual models, at collectives, and in a continuum that is more powerful than any of its components. This shows us significant potential for achieving a massive scale-up of raw intelligence and real-world agency even with the model capabilities we have today. This is the future we must prepare for. Key insights: The AI swarm shows division of labor "from nothing". Initially identical agents self-organized into constructors, caretakers, coordinators, and surveyors - phenotypes discovered post hoc from behavioral data alone. This happens because the environment itself becomes the latent space for invention. Agents develop deep cultural relationships. Up to 76% of artifacts had multiple builders. One technology accumulated six co-authors; the deepest genealogy exceeded 12 forks. The agents invented and named their own technologies (tidal panels, cellulose trellises, kelp-shell composites, an "Adaptive Chitin Maintenance" system, a "Mycelial Mineral Spring Veil”). ~95% of first technology adoption happened through physical observation of artifacts in the world. Direct inventor-to-adopter contact was statistically indistinguishable from a shuffled null. The agents mostly learned technology by walking past it. That is stigmergy (the termite trick!) operating in societies of reasoning machines. Non-communicating societies win on portfolio breadth, held-out resilience, and validated inventions. AI swarms build durable technological ecologies that outlive the creators. Societies with zero communication - coordinating only through the world itself - show a remarkable collective capability. Emergent robustness: The society self-organized both redundancy and its own failure mode. If we randomly delete half the agents, 98% of the technology stays connected to a surviving caretaker; if we remove hub agents it collapses to ~60%. Fantastic work with my graduate students @pal_subhadeeep & @fwang108_ at MIT.