State of AI
Q4 2026

Five things that defined the AI trade through Q3 2026 and will shape how Q4 plays out.

Zoom out from any single number and Q3 2026 reads as the quarter the AI trade’s two competing narratives — genuine, revenue-backed demand versus an increasingly stretched, financially engineered rally — both got real evidence behind them at the same time.

We flagged the bear case as it was forming. Back on June 30, when Michael Burry first disclosed his short book via a Substack post, we covered it in our look at what an AI bubble burst would mean for global markets: the semiconductor index was already up 102% year to date and trading 65% above its 200-day moving average — a level matched only once before, at the dot-com peak — while the top 10 AI stocks had gained 784% over twelve months, outpacing even 2000’s 622% run-up. That piece landed on a measured middle position: AI can be genuinely transformative and the stocks can still be overvalued at the same time. Everything that followed in July and August — the chip selloff, the sharper version of Burry’s depreciation argument — is that thesis playing out in real time rather than a new story.

The other half of the quarter’s story is what happens once OpenAI actually goes public. We covered the mechanics of that in our piece on Microsoft’s OpenAI stake and what it does to Microsoft’s earnings: Microsoft’s roughly 27% stake, then valued near $230 billion, means a 10% quarterly swing in OpenAI’s valuation moves Microsoft’s headline net income by an estimated $23 billion — noise that has nothing to do with Azure’s underlying growth. With OpenAI’s confidential IPO filing now in motion, that accounting-noise problem stops being hypothetical and starts showing up in Microsoft’s actual reported numbers, likely within the next two to three quarters.

Put together, Q3 was the quarter the abstract bubble debate got a set of concrete, dated events attached to it: a specific short book, a specific one-day selloff, and a specific IPO filing that will force real audited numbers into public view. Q4 is where those threads start resolving one way or the other.

The four largest hyperscalers are now guiding to a combined $725 billion in 2026 capital expenditure — up 77% from $410 billion in 2025. Microsoft raised its calendar-year guidance to $190 billion, well above the $152 billion analysts had modelled. Alphabet matched it at $190 billion after a mid-year $5 billion upward revision. Meta lifted its own full-year range by $10 billion, to a figure “topping $145 billion.”

What’s new this quarter isn’t the scale of the spending — it’s why some of it is going up. Microsoft CFO Amy Hood attributed $25 billion of the increase directly to rising memory chip and component costs. That is a meaningful tell: the AI buildout has grown large enough to move prices in its own input markets, which is exactly the dynamic behind the semiconductor stock selloff covered below.

If you wanted one number to argue the AI trade is still fundamentally sound, Nvidia handed it to you on August 26. Q2 FY2027 revenue came in at $96.2 billion, up 106% year over year. Data Center revenue — the number that actually matters for the AI story — was $89.0 billion, up 117%. Gross margin held at 75%. Guidance for Q3 came in at $108 billion, plus or minus 2%, and Nvidia said that figure excludes China compute revenue entirely, meaning the underlying demand picture in the rest of the world is even stronger than the headline suggests.

Huang’s framing matters because it’s a direct answer to the bear case: that AI infrastructure spending is running ahead of any actual revenue it generates. Nvidia’s numbers, taken at face value, say the opposite — that demand for compute is now outpacing even Nvidia’s ability to supply it, with a “golden age of new AI labs and startups” scaling frontier models in parallel rather than one company driving the whole cycle.

Nvidia’s own quarter was clean. The broader semiconductor complex has not had a clean few months. On July 28, chip stocks sold off sharply and simultaneously across three continents. In South Korea, SK Hynix fell more than 13% and Kioxia-linked names and Korean peers followed; Samsung Electronics dropped over 12%, LG Innotek fell 18%. In Japan, Tokyo Electron fell 11%, Advantest over 10%, SoftBank Group 6.3%. TSMC, usually the calmest name in the group, still slipped nearly 3%. The U.S. semiconductor ETF (SOXX) fell more than 2% on the same news.

The stated triggers were specific rather than vague macro nerves: broker notes warning of a “memory price peak in 2027,” and media coverage of China’s stated ambitions in memory chips and lithography equipment — a direct competitive threat to the Korean and Japanese suppliers that had been among 2026’s best-performing stocks. It was the first genuinely sharp, multi-day drawdown the AI hardware trade had seen since the rally began.

And the bear case now has a name attached to it

Michael Burry — of Big Short fame — has turned that one-day wobble into a standing argument. He is short Nvidia, has more than halved his Palantir position, and opened a new short against Micron, alongside Applied Materials and the SOXX ETF. His core technical claim is about depreciation: he argues hyperscalers are extending the useful accounting life of Nvidia chips and other AI computing equipment in a way that understates depreciation expense, and he estimates the resulting earnings overstatement could total roughly $176 billion between 2026 and 2028. He has separately flagged more than $1 trillion in what he calls “Byzantine” vendor-financing arrangements — deals where AI infrastructure suppliers help fund their own customers’ purchases — as obscuring how much of current “demand” is genuine end-user pull versus circular financing. He has compared current market sentiment to the final months of the 1999–2000 bubble and warned of a possible “1987-type” air pocket in the S&P 500.

While the money story dominated headlines, the model release cadence kept compressing. Anthropic shipped Claude Opus 5 on July 24, positioned as the reasoning ceiling for commercially available models and introducing an “Effort Dial” to modulate how much chain-of-thought compute a query uses; Claude Sonnet 5 had already landed as the enterprise default in late June. OpenAI answered with the GPT-5.6 family in early July — Sol at the frontier tier, Terra at roughly half the cost, and Luna, cut to $0.20 per million input tokens by the end of the month. Google shipped Gemini 3.7 Flash in mid-August with a native 2-million-token context window. On the open-weight side, Moonshot’s Kimi K3 (2.8 trillion parameters, Mixture-of-Experts) landed in mid-July, keeping Chinese labs at the frontier of open-weight releases.

The pattern from last quarter — a model race measured in weeks rather than years — has held. What’s new is how aggressively pricing is moving at the low end (GPT-5.6 Luna’s price cut is the clearest signal) even as the high end (Claude Opus 5, GPT-5.6 Sol) keeps pushing capability. That combination is exactly what you’d expect if usage is genuinely scaling rather than plateauing — it supports Huang’s “compute is revenue” framing more than it supports a slowdown story.

The clearest scoreboard change of the quarter: Anthropic’s annualized revenue run rate hit $65 billion at the end of July, up from $47 billion in May and just $9 billion at the end of 2025 — $18 billion of that added in two months alone. That run rate now sits well above OpenAI’s, which reached $40 billion, itself double its end-2025 figure. Investors are reportedly modelling Anthropic finishing 2026 somewhere between $100–120 billion in annualized revenue if the growth rate holds.

Both companies are now formally on the IPO path. Anthropic filed to go public around June 1; OpenAI followed with a confidential filing roughly a week later, saying only that the filing “gives us the option to go public sooner if that ends up being best” without committing to a date. Secondary-market pricing has moved with the revenue story: Anthropic near a $1 trillion valuation, up 123% year to date, versus OpenAI around $880 billion, up a comparatively modest 11.3%. Neither company has confirmed a listing date, and both are reportedly wary of the calendar — SpaceX is expected to list before either of them, potentially absorbing investor capital that would otherwise compete for their offerings.

Three things matter most over the next 90 days.

Watch #1 — Whether the Semiconductor Wobble Was a Blip or the Start of Something

One sharp session doesn’t make a trend, and chip stocks had already run enormously in 2026. But the specific trigger — a 2027 memory price peak call plus China’s stated memory ambitions — is a structural concern, not a sentiment one. Watch memory pricing data and any follow-through selling in SK Hynix, Samsung, Micron and the SOXX complex as the real signal, more than any single day’s move.

Watch #2 — The First IPO to Actually File

Anthropic and OpenAI have both filed; neither has set a date. If either moves from confidential filing to an actual roadshow in Q4, it becomes one of the largest technology listings ever attempted and gives the market its first real, audited look at frontier-AI unit economics — the single best real-world test of the “is this a bubble” question either side of the debate has.

Watch #3 — Whether Burry’s Depreciation Argument Shows Up in Q3 Earnings

Big Tech’s Q3 earnings (reported through October and November) will show updated depreciation schedules and capex guidance. If any hyperscaler quietly shortens the useful life it assigns to AI chips — effectively conceding Burry’s point — margins on paper get worse even if the underlying business doesn’t change. That’s a much more concrete thing to watch for than any single stock price move.

Data in this report draws from Nvidia’s Q2 FY2027 results release, CNBC coverage of the July 28 semiconductor selloff, reporting on Michael Burry’s short positions and depreciation thesis, TechCrunch on Anthropic’s revenue, TechCrunch on the OpenAI and Anthropic IPO filings, Big Tech Q2 2026 earnings reports and capex guidance, and public model release documentation from Anthropic, OpenAI and Google. All figures accurate as of September 1, 2026.

This report represents the editorial opinion of AllinAllSpace and does not constitute financial or investment advice. AllinAllSpace is not a registered investment advisor.