On the twelfth of November, Jensen Huang resigned as chief executive of Nvidia, and Colette Kress, the company’s chief financial officer, took over as interim head of the most valuable company on earth. For a week the story had been an affair and a boardroom mess. Then, on the thirteenth, the New York Times, CNBC and Al Jazeera put reporting behind it, and it became something far heavier. U.S. agencies had concluded that the woman at the centre of it — Alice Huang, a Megaspeed executive, no relation to Jensen — was aiding Chinese intelligence, and that her firm had been diverting banned Nvidia chips into China in violation of export controls.
Read it slowly, because the fairness matters. Intelligence also concluded that Jensen Huang himself did not know her role — a point we carry deliberately. What makes this a Washington story rather than a tabloid one is the timeline the Times reports: a secret memo in the spring of 2026, its drafting reportedly supervised by Secretary of State Marco Rubio, warned she was a spy for China’s Ministry of State Security; the President was briefed in late 2025 and chose not to act; and the federal case moved from Commerce to the Justice Department this month, after an earlier Commerce probe was reportedly halted when Nvidia’s own lawyer intervened. Senator Ron Wyden and Representative Richard Neal have opened an investigation into how the administration handled it, and Huang has agreed to testify to Congress before Thanksgiving.
The honest caveat, because we insist on it: none of the espionage allegations have been adjudicated, the figures around Megaspeed’s chip purchases are secondary and contested, she denies wrongdoing, and the Justice Department says it has reached no conclusions. But strip the scandal away and the structural fact is the one this index keeps returning to: the entire AI build rests on a single company’s chips, that company has just lost the man who was its whole strategy, and its first quarterly results without him land this week under a chief executive doing the job for the first time.
Two weeks after filing, Anthropic set the terms: seventy to seventy-eight dollars a share, a fully diluted valuation of roughly one-point-four trillion dollars, and as much as sixty-four billion dollars raised — a roadshow that began the week of the ninth and a debut expected before Thanksgiving. The number is staggering and, in context, a climbdown: only weeks ago bankers were reportedly floating two trillion. The market is still willing to buy the AI story, in other words — just not at the price it was dreaming about in October.
This is the cleanest test we have of what public investors actually think a model-maker is worth, and it lands in the same fortnight the stock market convulsed over AI spending and Nvidia lost its chief. Reporting since suggests the order book is heavily oversubscribed, which cuts against the simple bubble-popping read: a 1.4-trillion-dollar company losing money today is a bet that the revenue curve eventually climbs to meet the spending. Disclosure, as always: we are an Anthropic customer and this index is written with their model, so weigh us accordingly.
SoftBank reported a seven-billion-dollar quarterly net loss, and the cause is almost poetically on-theme: a three-point-seven-billion-dollar writedown on data-centre power equipment, taken after its US plans changed and left it holding gear it had already paid for in advance. Financing costs more than doubled over the quarter. The company noted that some future purchases could still qualify for up to two-point-one billion dollars in rebates — a silver lining on a very expensive cloud.
This is the most tangible version of the week’s whole argument. The AI build-out is usually described in commitments and forecasts; here it turned into a realised loss on real equipment, bought ahead of plans that then shifted. It is the cost of moving first in a market where the ground keeps moving — and a reminder that “capex” is not an abstraction. It is machines someone has already paid for, whether or not the building meant to hold them ever gets its permit.
Underneath the headlines, the plumbing tightened. Bloomberg reported that Blackstone’s QTS, one of the largest data-centre operators, chose to raise a smaller private-placement deal rather than borrow a much larger sum at today’s higher cost — a refinancing that resizes rather than defaults, but a telling choice all the same. Separately, the Banker reported that lenders are increasingly using credit insurance to offload part of their data-centre exposure and free up room to keep lending.
Neither is a crisis, and we won’t dress them as one — the QTS figure is single-sourced, and a smaller deal is not a failed one. But together they sketch, from the financing side, the same picture SoftBank sketched from the spending side: the cheap, limitless money that underwrote the first phase of the build is getting choosier. And the honest worry with the insurance trick is that it does not remove the risk, it moves it — from the banks that understand these loans to the insurers who may not.
On the twelfth, a coalition of non-profits filed suit against Anthropic, Google, Meta, OpenAI and SpaceX’s AI division, asking a court to halt the development of more advanced systems on the grounds that they pose catastrophic risks — the plaintiffs point specifically to bioweapons and cyberattacks. It is, to be clear, a request for a court to intervene, not an order that stops anything; the odds of a judge freezing the US AI industry by injunction are long.
What earns it the slot is not the likelihood of winning — it’s that the catastrophic-risk argument has graduated from open letters and conference panels into a filed legal complaint naming the five biggest builders at once. Whatever the court does with it, that is a new front. The near-term business consequence is uncertain and probably small; the signal — that “slow down” now has a docket number — is not.
$2.88, up twenty-seven cents — and once again the move is composition, not a market verdict. One thing changed. Google’s new premium flagship, Gemini 4 Argon, cleared its four-week general-availability window and rolled into the basket in place of Gemini 3.6 Flash — at $8.80 against $3.30, a premium model displacing a mid-tier one, which lifts the geometric mean almost on its own. Every other posted list price held. The irony runs the other way this fortnight: the posted price of frontier AI ticked up, in the same days the market lost Nvidia’s founder and spent itself arguing the whole thing costs too much to build.
The structural fact underneath still holds. The cheap tier — two models at twenty-two cents — is where the floor keeps dropping, and Chinese open-weight models still carry more than half of global token consumption. The index rose this week only because a dearer flagship entered, not because anyone’s bill went up. Cheaper to run every week; the entire argument is still about what it costs to build.
| Model | Provider | Blended $/M | LN(blended) | vs Ed.022 |
|---|---|---|---|---|
| FrontierClaude Fable 5.1 | Anthropic | $22.00 | 3.091 | ▶ flat |
| PremiumClaude Opus 5.5 | Anthropic | $8.80 | 2.175 | ▶ flat |
| MidGPT-5.6 Terra | OpenAI | $5.00 | 1.609 | ▶ held |
| MidGPT-6.1 Sol | OpenAI | $4.40 | 1.482 | ▶ flat |
| MidClaude Sonnet 5 ❅ | Anthropic | $4.40 | 1.482 | ❅ frozen |
| PremiumGemini 4 Argon | $8.80 | 2.175 | ▲ new (in for Flash) | |
| CheapQwen3.7 Max | Alibaba | $2.64 | 0.971 | ▶ flat |
| CheapClaude Haiku 5.5 | Anthropic | $0.22 | -1.514 | ▶ flat |
| CheapKimi K2.6 (open) | Moonshot AI | $1.87 | 0.626 | ▶ held |
| CheapGPT-6 Luna | OpenAI | $0.22 | -1.514 | ▶ flat |
Gemini 4 Argon went to general availability on 15 October and has now cleared the standing four-week window, so it rolls into the basket in place of Gemini 3.6 Flash — a premium flagship displacing a mid-tier model, and the whole of this edition’s twenty-seven-cent rise. The rest of the recent wave stays under the same test, with prices ready for the roll the moment each clears. We report every decision as it lands, rather than churning the basket on launch-week noise.
| Model | Provider | Input $/M | Output $/M | Blended $/M | Status |
|---|---|---|---|---|---|
| Gemini 4 Argon | $4.00 | $20.00 | $8.80 | ✓ Actioned — rolls in for Gemini 3.6 Flash | |
| GPT-6.1 Terra | OpenAI | $1.00 | $5.00 | $2.20 | Under review (clears ~late-Nov) |
| Kimi K3.1 Pro | Moonshot AI | $2.00 | $10.00 | $4.40 | Under review (clears ~late-Nov) |
| GPT-6 Astra | OpenAI | $10.00 | $50.00 | $22.00 | Under review (full GA) |
This fortnight’s money story — the founder of the company the whole build rests on gone and about $1.4 trillion off Nvidia since its October record, four central banks warning about the same thing in a single week, and one IPO filing putting a public price on the trade — is exactly what the price table alone cannot show. AI Vital Signs is our standing, weekly-updated dashboard of the money behind AI: the total capital at risk, the tower of debt and off-balance-sheet commitments, US versus China, the labs’ revenue against what they have promised to spend, and the circular financing that ties it all together. New on the board this fortnight: the Federal Reserve’s Vice Chair for Supervision putting AI-related debt “on track to exceed $800 billion this year” with off-balance-sheet commitments estimated at “more than three times” that, and naming “circular financing” outright — echoed by the ECB, the Bank of England and the RBA in the same week; Anthropic’s $70–78 IPO range (~$1.4 trillion fully diluted), with $26 billion of stock bought by its own chip and cloud suppliers; SoftBank’s leverage now partly disclosed — group loan-to-value 15.9%, net asset value down about a quarter in six weeks, a $3.7 billion writedown on data-centre power gear; and OpenAI’s first close of about $11 billion at a $1.25 trillion valuation, roughly a third of its original target.
Open AI Vital Signs →A curated log of significant workforce reductions where AI was cited as a factor, and documented reversals where roles cut for AI reasons were subsequently reinstated. Updated each edition.
| Company | Headcount | AI cited | Date | Context |
|---|---|---|---|---|
| Amazon | ~14,000 | ✓ Yes | Oct 2026 | Corporate roles cut — the company's reporting tied it directly to AI “taking over routine work.” The largest single AI-cited corporate cut of the year, and Amazon alone now accounts for roughly half of 2026's tech layoffs. Reuters; NewsNation. |
| Canva (AU) | ~1,271 | ~ Partial | 4 Nov 2026 | The Melbourne company's first-ever major restructure, framed by its CEO as a reset after its “hardest year yet.” The honest read: the Financial Review reports the cuts largely target AI-related roles rather than AI replacing workers — a reshuffle around AI, not a clean replacement story. Forbes Australia; HCA Mag; ACS. |
| Barclays (UK) | ~5,000 (~5%) | ~ Partial | 22 Oct 2026 | A bank cutting about five per cent of its staff in a cost drive. The honest read is in the CEO's own framing: he called AI the “icing on the cake” — the restructuring is the cake, AI the accelerant. The clearest sign yet that banks are starting to bank the efficiency. ITV; Reuters via Business Day; FinTech Futures. |
| Company | Roles cut | Status | Date | What happened |
|---|---|---|---|---|
| Commonwealth Bank (AU) | ~40–45 | ✓ Rehired | 2026 | AI voice bot deployed as replacement generated a surge in call volume rather than reducing it; roles reinstated; CBA issued a formal acknowledgement. The documented case that started this tracker. |
Short-form video essays on AI in business — the thinking behind this index. Evidence, not hype.
The Index is the geometric mean of the ten member models' blended prices, expressed in USD per million tokens. The geometric mean treats proportional differences equally — a model twice as expensive as another carries the same weight whether the comparison is $0.20 vs $0.40 or $20 vs $40. This produces a more balanced picture than an arithmetic mean, which is dominated by the most expensive entries.
The blended price for each model is input × 0.7 + output × 0.3. This is the same formula as the public Token Price Index. Cache hit pricing is excluded because cache-hit rate is a workload assumption — including it would mean the Index measures workload patterns as much as posted prices. Temporary promotional pricing is also excluded: the basket uses standing list prices, so a model's entry reflects a durable posted price rather than a limited-time offer. The Index is a posted-price market reference, not a forecast of any organisation's actual bill.
The Index rises when the market composition shifts toward more expensive models, or when individual member prices rise. It falls when cheaper models enter or when prices fall. Mix changes are intentional — the Index tracks the frontier as the market defines it, not a fixed historical basket. A flat edition means every posted list price held and the basket was unchanged; this edition was not flat, because Google’s Gemini 4 Argon ($8.80 blended) rolled into the basket in place of Gemini 3.6 Flash ($3.30) after clearing the four-week general-availability window, lifting the Index about twenty-seven cents even as every other list price held.
Posted prices are not realised costs. Organisations typically pay less because of enterprise discounts, prompt caching, model routing, batch processing, and committed infrastructure spend. The Index is the published list price; a separate practitioner calculation incorporating cache hits and other discounts can run 50–70% lower for production workloads with prompt caching configured.
The Innovare Index is curated to ten models drawn from at least five providers, with at least one entry in each capability tier (Frontier, Premium, Mid, Cheap). To enter the Index, a model must meet all six criteria: