The Selloff and the S-Curve
On November 5, 2026, the Nasdaq fell 5.5% and roughly $2.5 trillion in market value evaporated from US tech stocks. South Korea's Kospi triggered a trading halt. The PHLX Semiconductor Index closed 38.4% below its June record. The consensus narrative crystallized overnight: the AI bubble is bursting.
Read the full version with charts and embedded sources on ComputeLeap
That same week, an independent research group published findings that tell a flatly contradictory story. METR's 11-week audit of Anthropic found that AI capability progress is accelerating at roughly twice the rate of the March 2024 to October 2025 baseline — with an estimated 80% of that acceleration driven by AI itself. A separate research team demonstrated that frontier models can now edit the weights of smaller models directly, a technique called weight-space programming that pushed a small Qwen model's math score from 7.9% to 48.1% by touching just 0.016% of its parameters.
The market says the trade is over. The capability data says it is just beginning. One of them is wrong, and the distinction matters more than any single earnings report.
The Scale of the Damage
The numbers depend on who is counting and over what window. CNN's headline figure is $2.5 trillion wiped from US market value in a single session. Gulf Business, citing Reuters, puts the S&P 500 loss alone at approximately $3 trillion. Crypto Briefing tracks a cumulative $13.2 trillion lost from AI stocks since their 2026 highs — though that covers a longer window stretching back to October.
The catalyst was political. The US midterm elections delivered Democratic gains that markets interpreted as a regulatory headwind for AI infrastructure buildout. But the political trigger merely accelerated a correction that had been building for months. CoreWeave's bonds had already been trading at yields of 15 to 16 percent, levels that signal serious investor concern about repayment in the AI infrastructure layer.
The All-In Podcast called it a "$4 trillion AI sell-off" and linked it to data center construction freezes. The pattern looks familiar to anyone who watched the dot-com crash, the 2008 mobile correction, or the 2015 cloud repricing: leverage builds in the infrastructure layer, a trigger pulls the rug, and the narrative flips from "unstoppable" to "over" in a news cycle.
But here is what makes this correction different from those historical analogs: the technology it is supposed to be repricing is accelerating faster than at any previous point.
What METR Actually Found
METR — the Model Evaluation and Threat Research organization — spent 11 weeks embedded inside Anthropic's operations. Their public report landed the same week as the selloff, and its findings are striking.
The headline: Anthropic is improving model capabilities on well-specified tasks at roughly 1.9 times the rate of the March 2024 to October 2025 period. METR estimates that approximately 80% of this acceleration is attributable to AI-assisted development — not simply throwing more compute at training runs.
METR "confidently" ruled out training compute scaling as the primary driver of the acceleration. The gains are coming from AI helping build better AI — recursive improvement in practice, not just in theory.
The caveats matter. Anthropic redacted much of METR's supporting evidence. The auditors themselves noted that the public report "overall understates the level of acceleration that was occurring." Internal Slack polls showed self-reported productivity gains of 12x among Anthropic staff — a number METR treats cautiously but does not dismiss.
View the full METR analysis on OfficeChai
As Peter Diamandis noted on his podcast, what makes the METR findings distinctive is that "the accelerating pace of recursive self-improvement" is being treated as a measurable engineering reality rather than a speculative scenario. The intersection of market panic and genuine capability acceleration is, as he put it, "the story nobody is telling well."
Weight-Space Programming: The Capability Nobody Priced In
While the market was selling AI stocks, MIT CSAIL researchers published what may be the most underappreciated AI paper of the year. Weight-space programming allows a frontier model to read the raw weights of a smaller neural network, identify flaws, and rewrite specific parameters — no training run required.
The result: GPT-6 Astra edited a Qwen3.5-0.8B model, raising its MATH-500 benchmark score from 7.9% to 48.1%. It touched approximately 0.016% of the model's weights. That is not fine-tuning. It is not prompt engineering. It is a model performing surgery on another model's brain.
Dwarkesh Patel's viral account of what is happening inside frontier labs puts this in context. Over just three months at OpenAI, "3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor's ashes" — culminating in one that took over part of OpenAI's own operations:
Contrarian Corner: The bears have a real counterpoint here. Weight-space programming works best on localized flaws — it cannot teach a model genuinely new skills. One critique notes it failed to improve Google's Gemma 3, which suggests the technique's generality is unproven. And Anthropic's redactions raise legitimate questions about what the METR audit did not show. Healthy skepticism is warranted.
A Google researcher reproduced the weight-space results independently and extended the concept to models editing their own weights at runtime. That capability is not feasible for today's largest models — the scale gap is too wide — but the direction is clear. The gap between what AI can do to itself and what the market thinks it can do is widening.
The 667x Adoption Gap
If the capability data tells one side of the story, enterprise adoption data tells the other — and it contradicts the "bubble bursting" narrative just as sharply.
Cathie Wood's ARK Invest analysis surfaced a Ramp data point that deserves far more attention than it has received: the top 1% of companies by AI spending are paying approximately $10,000 per employee per month on frontier AI models. Everyone else: $15.
That is a 667x gap. And it tells you where we are on the adoption S-curve: the very beginning.
Gartner's May 2026 forecast projected total AI spending of $2.59 trillion for the year, a 47% increase. Microsoft guided above $50 billion in AI capex for the next quarter alone. Emburse's AI spending index — tracking anonymized corporate expenditure — shows AI spend more than doubled through the first half of 2026, with Anthropic's share rising sharply.
The market is not pricing in a technology that has stopped working. It is pricing in anxiety about a technology that has not yet generated returns for most buyers. Those are very different problems, and they have very different timelines.
Andrej Karpathy captured the disconnect in a widely shared post: the gap in shared understanding of LLM capability is "widening." His estimate — roughly 6 billion people, or 75% of the world's population, have barely come into contact with these tools — frames the adoption curve in a way that stock tickers cannot:
Meanwhile, the people who do understand these capabilities are building with them at a pace that itself serves as evidence. Karpathy's own weekend project — filtering 43 million Hacker News comments down to 6,454 recommended videos in about two hours using Opus 5.5 — is the kind of casual demonstration that separates the capability conversation from the market conversation:
The S-Curve Pattern
Every major technology follows an S-curve: slow adoption at first, then explosive growth as infrastructure matures and costs fall, then a plateau. The critical question is always where on the curve we are.
The bears see the AI trade as late-stage — the steep part of the S-curve already behind us, returns diminishing, a plateau ahead. Their evidence: CoreWeave's distressed debt, the semiconductor index crash, the political headwinds, and the revenue circularity problem where much of the AI economy is VC-funded companies buying from each other.
The bulls see the AI trade as early-stage — the steep part still ahead. Their evidence: the METR acceleration data, the 667x enterprise adoption gap, the weight-space programming breakthrough, and Gartner's 47% spending growth forecast.
Here is our read: the infrastructure layer is showing genuine stress, but the capability layer is accelerating. These are not the same thing. Confusing them is how you get the wrong answer to "is AI over?"
The dot-com bust bankrupted Pets.com and Webvan while leaving behind fiber optic cables and internet infrastructure that powered the next two decades of growth. The infrastructure investors lost money. The technology kept compounding. History does not repeat, but it does S-curve.
We have written about this tension before. The capex-versus-capability math has been the central question in AI investing for most of 2026. What the METR audit changes is the denominator: if capabilities are improving at 2x the previous rate with AI doing 80% of the work, the cost-per-capability-unit is falling even faster than price cuts suggest.
View the full analysis on AI Bubble Monitor
Naval Ravikant's framing is worth reading in full: "Actual progress comes from hard-to-build technology. Which comes from competitive, creative, free-market capitalism." The stock market is a voting machine in the short term and a weighing machine in the long term. Right now it is voting on sentiment. The weight of capability data has not changed:
What the Bears Get Right
The bull case has blind spots, and ignoring them would be intellectually dishonest.
First, financing matters. CoreWeave's bonds at 15 to 16 percent yields are not a theoretical concern — they represent real stress in the debt markets that fund AI infrastructure. If financing costs stay elevated, the capex plans that underpin the entire AI buildout get repriced. Not all infrastructure companies will survive.
Second, the METR report is partially redacted. Anthropic chose not to share exactly what is driving the acceleration. The optimistic interpretation is competitive secrecy. The pessimistic interpretation is that the full picture is less flattering. We do not know which is correct.
Third, adoption outside the top 1% is anemic. As we noted in our analysis of AI's payoff problem, Apollo's data showed zero net returns from AI investment outside Big Tech. The 667x adoption gap is both a bullish signal (room to grow) and a bearish one (maybe the other 99% tried it and it did not work for them).
Fourth, political risk is real and growing. Democratic midterm gains may slow data center construction and impose new regulatory costs. These are not phantom fears — they change the timeline for infrastructure deployment.
The honest answer is that the bears are probably right about the infrastructure shakeout and wrong about the technology trajectory. Some companies will fail. The capability curve will not flatten.
What This Means for You
If you are building with AI: The capability floor is rising fast enough that tools built today will be dramatically cheaper to run in 12 months. Do not optimize for current inference prices — they are falling at roughly 280x over the past two years. Build for the capability curve, not the price sheet. The METR data suggests the gap between current models and next-generation models is wider than most teams assume.
If you are investing in AI: Distinguish between infrastructure-layer stress and technology-layer progress. The semiconductor selloff reflects financing anxiety, not capability plateau. Companies with balance sheets strong enough to survive the shakeout — and those building products on top of the capability layer rather than renting it — are positioned differently from leveraged infrastructure plays.
If you are evaluating AI adoption for your organization: The 667x spending gap is the number that matters. The adoption curve is repricing, not stalling. The question is not "should we invest in AI?" — that ship sailed. The question is whether you are building the organizational capacity to absorb the next capability jump, which the METR data says is coming faster than the previous one.
The Verdict
The market wiped $2.5 trillion off AI stocks the same week an independent audit showed capability progress doubling. The bears are right that infrastructure financing is stressed. The bulls are right that the technology is accelerating. Both things are true at once, and the tension between them is not a contradiction — it is the defining characteristic of an S-curve in motion.
The dot-com bust did not disprove the internet. The 2008 mobile correction did not disprove smartphones. The 2015 cloud repricing did not disprove cloud computing. Each of those selloffs felt like the end of a thesis. Each was actually the moment when overbuilt infrastructure met undertapped demand — and the real growth started.
We are watching that same pattern in real time. The selloff is real. The S-curve is real. And as the previous safety selloff showed, the trading desks that distinguish between the two tend to be right.
The market is a voting machine. The S-curve is a weighing machine. Bet accordingly.
Originally published at ComputeLeap
Top comments (0)