AI Bubble Tracker

“The dependency of the hyperscalers on Anthropic and OpenAI is just huge — both lose billions and rely on raising capital for survival. Oracle's debt is rated barely above junk. There is no hedge for the LLMs: hyperscalers have fallback franchises, the labs don't.”

“First full read across all four hyperscalers: Microsoft's commercial backlog jumped to $678B (from $625B), closing in on Oracle's $638B. Google Cloud posted $155B (+82% YoY) and Amazon's AWS posted ~$200B — both reported here for the first time. Backlog growth is not just an Oracle story anymore; it's broad-based across the sector.”

“No fresh backlog figures this week — Microsoft, Google and Amazon already reported their quarters, and Oracle's next earnings (fiscal Q1 FY27) isn't until Sept 9-12, which is this signal's next real checkpoint.”

“Still no new backlog print — Oracle has now confirmed its fiscal Q1 FY27 earnings date as Thursday, Sept 10 (after market close), so the number itself remains four days away as of this sweep. The wait is happening against a noisier credit backdrop: Oracle shares have continued sliding (down roughly 30% over the trailing month per multiple trackers) on investor worry about its debt load and its concentration in a single, unprofitable customer (OpenAI is reported to account for over half of the $638B backlog). None of that changes the backlog NUMBER, but it raises the stakes on whether Sept 10's print shows growth continuing or cracking.”

“Oracle reported Sept 10: backlog hit a record $664B, up $209B year-over-year — but the QUARTER-OVER-QUARTER pace slowed sharply, from about $85B of new backlog added last quarter to about $26B this quarter (roughly 46% YoY growth this print, down from the ~363% YoY comparison logged a quarter ago). That is the first real sign of deceleration this signal has shown since this tracker started — moving status from 'accelerating' to 'watch' until next quarter's report (expected around December) either confirms a genuine slowdown or shows it bouncing back. Oracle's own capex jumped to $28.5B for the quarter (from $8.5B a year ago) and it raised $20B in a stock sale, both signs it's still spending hard against that backlog. The stock popped as much as 8.5% right after the report, then reversed to close down about 1.7% — investors cheered the headline number, then had second thoughts. No update from Microsoft, Google, or Amazon this week (their next earnings are late October).”

“Nvidia's $500B securitized-financing push means some data-center capex shifts off hyperscaler cash flow — hyperscaler FCF 'could improve' optically while the debt stack extends into banks and private credit. Read improving FCF with suspicion from here.”

“Kill-trigger watch candidate: Meta is getting squeezed — latest quarter revenue +28% but expenses +55%, no cloud business to monetize AI through. Most likely first major player to cut; watch how the market reacts when it does.”

“The five biggest spenders (Amazon, Alphabet, Microsoft, Meta, Oracle) now guide to a combined $775-800B for 2026 — roughly 3x the ~$238B spent in 2024. This is the opposite of the kill trigger so far: Alphabet RAISED its 2026 capex guidance (to $195-205B) and its stock fell about 7% over the following week — a capex increase punished, not a capex cut rewarded. That's still consistent with the current framing (investors want the spending to keep climbing), just a reminder the market is getting twitchier about how big the numbers are getting.”

“Still no kill trigger, and the numbers keep climbing: Nvidia's Aug 26 earnings (record $96.2B quarterly revenue, Data Center alone $89B, Q3 guided to $108B) had its CFO reiterate that the top hyperscalers' combined capex is heading toward roughly $1.3T next year, up from about $800B now. But the same 'raise-without-proof gets punished' pattern repeated: Meta fell about 10% on a JPMorgan downgrade after guiding 2026 capex up to $125-145B, the second capex raise in three weeks (after Alphabet) to get sold off rather than bought. Microsoft and Amazon, whose capex raises are tied to demonstrated cloud revenue growth, are not seeing the same reaction — the market is starting to price the difference between spending that's paying for itself and spending that isn't.”

“No new capex guidance this week — none of the five biggest spenders reported earnings (the next round lands in October for Q3). Still no kill trigger. The clearest sign spending keeps translating into real infrastructure: OpenAI said its new GPT-6 Astra model, released Sept 3-4, was trained on more than 100,000 GPUs at its Stargate site in Texas — by far its largest training run — which is capex showing up as product rather than as a guidance number. Separately, investor Michael Burry kept up his public campaign arguing the whole buildout rests on circular financing, this time estimating roughly $879B in hyperscaler commitments circling through Nvidia; that's a credibility challenge to the spending story, not evidence of an actual cutback.”

“Oracle's Sept 10 earnings were this week's only real capex data point: it spent $28.5B in the quarter (versus $8.5B a year ago) and held its full-year guidance flat at $90-95B gross capex — not a cut. Still no kill trigger anywhere: nobody announced a spending cut, let alone had their stock rewarded for one. Oracle's stock actually did the opposite of the kill-trigger pattern — it rallied on the spending/backlog news, then gave the gain back within the same session, which reads more like investors getting nervous about HOW the spending is funded than about the spending itself.”

“Red here does not mean Chinese models are failing — it means pressure against US share is RISING. This signal measures the force pushing on US revenue speed.”

“Independent corroboration: enterprises are using Chinese open-weight models more and more; 'overcharging for tokens is going to be very difficult with good-enough competition.' A price war is possible — and Anthropic + OpenAI are ~70% of hyperscaler AI revenue, so a price war transmits straight into the backlog.”

“New Chinese releases kept coming this week (Alibaba's Qwen3.8 Max Aug 2, Z.AI's GLM-5.2 Turbo Aug 17), but the readings on actual Chinese-model traffic share are noisy right now — trackers show anywhere from ~45% to over 60% of OpenRouter volume depending on whether you measure the top 10 models or all traffic. Treat the exact percentage as uncertain this week; the direction (more competitive Chinese releases, more enterprises adopting them) is not.”

“More Chinese releases (Z.AI's GLM-5.3-Flash Aug 26) and readings still cluster high, 51-61% of OpenRouter traffic by some trackers, with DeepSeek the single largest provider at roughly 16%. One genuine crack in the 'nearly free' narrative: DeepSeek's new V4 Pro model raised prices roughly 14x with new peak/off-peak billing — the first sign a leading Chinese lab thinks it can charge more, not less. Moonshot's Kimi K3 reportedly had to pause new subscriptions due to demand. Keep treating the precise percentage as noisy; the competitive pressure itself is not in question.”

“Alibaba's Sept 2 refresh of Qwen3.8-Max took the #1 spot on a major coding leaderboard while leaving pricing untouched — the opposite of DeepSeek's recent price hike, and a reminder that the 'nearly as good, nearly free' pressure this signal tracks hasn't gone away just because one Chinese lab tested higher prices. On the US side, OpenAI answered competitively with GPT-6 Astra (Sept 3-4), its most capable model yet — the two labs are clearly racing each other as much as they're racing the cost pressure from China. Exact OpenRouter share numbers remain too noisy across trackers to log a new reading this week; the direction of competitive pressure is unchanged.”

“Quiet week for new Chinese model releases — nothing dated this week from Moonshot, DeepSeek, Alibaba, or Z.AI. One useful data point resurfaced: a Sept 1 snapshot puts Chinese-origin models at roughly 46% of identified enterprise token usage (a different measure than the OpenRouter US-vs-China split tracked above, so not logged as a new point on that same series, but directionally consistent with 'pressure keeps building'). On pricing, OpenAI's new GPT-6 Astra actually launched ABOVE its prior model's price ($10/$50 per million tokens) rather than cutting — worth noting as a case where the US labs are charging a premium for their best model instead of racing China to the bottom, even as the volume-share pressure keeps building underneath.”

“Both labs are still accelerating, not just growing. Anthropic told investors its annualized revenue run-rate hit $65B in July — up 7x from a year earlier — and investors expect it to keep growing at roughly that pace, finishing 2026 around $100-120B. OpenAI's run-rate reached about $40B, roughly double where it stood in the spring. This is the strongest reading this signal has had — moved from 'watch' to 'accelerating' this week — but both companies are still deeply unprofitable, so the speed is what's propping up their valuations, not earnings.”

“No fresh run-rate print this week — largely a restatement with texture: Anthropic's preliminary Q2 revenue came in above $11.5B versus $787M a year earlier, and OpenAI's enterprise revenue passed its consumer revenue for the first time in July, with its paying customer count up 32% in a single month. Nothing here suggests deceleration; there's just no brand-new number to move the needle beyond what was already captured last week.”

“No new run-rate figure this week, but a genuine new revenue-generating surface: OpenAI rolled out GPT-6 Astra (Sept 3-4) — its most capable model yet, described internally as its largest training run ever — to ChatGPT Plus/Pro/Business/Enterprise, the API, and AWS, with API pricing set at $10/$50 per million input/output tokens. It's too soon for this to show up in a run-rate number, but a materially better product landing broadly is exactly the kind of thing that would show up in next month's read. Nothing suggests deceleration.”

“Quiet week — no fresh run-rate or revenue disclosure found from either company. Carrying forward Anthropic's ~$65B annualized run-rate (July) and OpenAI's ~$40B (August). Nothing found this week points toward deceleration.”

“The deal now has names attached: Nvidia formalized memorandums of understanding with six of the biggest names in finance — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — to raise the $500B+ in compute financing. Two new details sharpen the risk picture: the SEC issued guidance in July that removes a risk-retention rule which would otherwise have required sponsors to keep a slice of this debt on their own books, and Nvidia has personally guaranteed (with its own money) that the chips used as collateral will hold their value. Both moves make it easier for capital to flow into these deals and easier for the risk to end up somewhere other than where it was created — exactly the 2006 dynamic this signal is watching for.”

“Still MOUs, not signed final agreements, and still no remittance or delinquency data publishing. But the SEC guidance behind the deal (dated Aug 17 in wider coverage) got more scrutiny this week: it explicitly takes sponsors 'off the hook' on certain look-through requirements for data-center securitizations. BlackRock's Larry Fink compared the moment to the birth of the mortgage-backed-securities market in the 1970s — which, read straight, is either a compliment about unlocking capital or an omen, depending on how this decade's version ends.”

“Still MOUs only — no report found of the Nvidia/Apollo/BlackRock/Blackstone/Brookfield/Goldman/KKR arrangements converting to signed final agreements, and still no remittance or delinquency data publishing. The main new development is louder skepticism, not new structure: investor Michael Burry escalated his public case (including a Sept 4 post) that hyperscaler AI spending runs through a circular web of vendor financing, this time estimating that roughly $879B in hyperscaler commitments circle through Nvidia in ways that can inflate reported revenue without necessarily reflecting new end-customer demand. Nvidia's credit-default-swap spread has stayed elevated on this kind of scrutiny. None of this is new hard data on the securitization deal itself, but it's the same underlying concern this signal is built to watch — who ends up holding the risk when the labs paying the bills aren't profitable.”

“Still MOUs only — no signed final agreements and no remittance/delinquency data found. Burry himself made a move worth noting: on Sept 10 he disclosed he had trimmed his overall bearish positioning and let his December-2026 Nvidia and Palantir put options expire rather than rolling them into new contracts, while keeping longer-dated (2027) puts on Palantir and the Nasdaq (QQQ). He kept his three biggest short positions in place — Oracle, Palantir, and Nebius Group — so this reads as reducing near-term bet size heading into a heavy earnings/auction week, not as him changing his mind about the thesis.”

“No movement this week — H.R.4142 / S.2177 remain in committee with no markup or floor vote found, and no other new Chinese-AI restriction bill or executive action surfaced. Ladder holds at DRAFT_LEGISLATION.”

“Still quiet — no committee markup, floor vote, or new executive action found on H.R.4142 / S.2177 this week. Ladder holds at DRAFT_LEGISLATION.”

“No committee markup, floor vote, or new executive action found this week either. One caveat: Congress.gov's live action log couldn't be directly checked this week (a network restriction blocked the fetch), so this is a search-based 'no news found' rather than a confirmed 'nothing happened' — worth a direct Congress.gov check next sweep. Ladder holds at DRAFT_LEGISLATION.”

“Reconfirmed this week at the same ~40-41% level — no material change. Still record-high concentration, just not moving faster or slower than before.”

“Eased to about 37.6% as of Aug 28, down from ~40-41%. Still within the range this measure has bounced around in recent months and still historically extreme concentration — not read as a material de-concentration of the index.”

“Ticked back up to about 38% as of Sept 1, from 37.6% the week before. This is still just normal week-to-week wobble within the same historically extreme range this measure has held for months — record-high concentration, not a new trend in either direction.”

“No clearly-dated new reading this week — figures found still cluster around 37-38%, the same range logged for the past month. Nvidia alone was cited at roughly 8.2% weight as of Sept 10. Holding at 38% pending a fresher print.”

“First concrete auction data point for this signal: the Aug 13 30-year bond auction ($31.3B) drew a bid-to-cover of 2.39 — weaker than the trailing 12-month average of 2.43 — at a high yield of 5.216%, the highest 30-year yield in 25 years. Demand wasn't a failure, but it was softer than normal at a moment when the government needs buyers to keep showing up. Worth watching whether this is noise or the start of a trend.”

“The exact mechanism this signal watches for showed up live: a weak 20-year Treasury auction in mid-August briefly pushed 30-year yields above 5.08% and triggered a sharp same-day selloff in AI-adjacent stocks (CoreWeave -8.3%, Tempus AI -6.9%) before the Treasury's doubled bond buybacks calmed the market back down. Then, on Aug 28, incoming Fed Chair Kevin Warsh used his first Jackson Hole speech to say rate hikes may still be needed given elevated inflation — a hawkish surprise that pushed 2-year yields from 4.22% to 4.30% and complicates the idea that buybacks alone can keep the government's own borrowing costs pinned down. Two genuinely new threads for this signal, both pointing the same direction: rising long-term borrowing costs are now visibly colliding with AI-linked asset prices.”

“A genuinely quiet week for new auction data — no Warsh follow-up speech found since Jackson Hole, and no fresh bid-to-cover print. The next real data points are imminent: Treasury has scheduled a 10-year note reopening for Sept 9 and a 3-year note auction for Sept 10, both announced Sept 3 and settling Sept 15 — those results, plus how AI-linked stocks react around them, are this signal's next checkpoint.”

“The Sept 9 10-year note reopening came in strong, not weak: bid-to-cover of 2.71 (the best of the last 28 nominal 10-year auctions since June 2024) at a yield of 4.834%, with indirect (largely foreign) bidders taking 79.2% versus a 72.7% average — real demand for US debt showed up when it mattered. The Sept 10 3-year note auction result could not be confirmed this week. Despite the strong auction, the 10-year yield itself kept climbing, reaching roughly 4.97-4.98% by Sept 11 — the highest since October 2023 — after a hotter-than-expected inflation print pushed the market's odds of a Sept 15-16 Fed rate hike up sharply. Read together: buyers are showing up fine, but the price they're demanding (the yield) keeps rising anyway, which is the more important half of this signal's story than the auction mechanics.”