ASI Brief

AI brief, Sunday 1 November 2026

Amazon's 10-Q answers the question its press release left open. It carries its Anthropic position at about $239bn against roughly $18bn invested, put $50bn into OpenAI during 2026, and has a $15bn facility that Anthropic can draw on as Amazon delivers compute. Mistral released open weights for a trillion-parameter model with a licence carve-out for hosted resale. And a company selling AI compute filed a prospectus disclosing substantial doubt it can continue as a going concern.

Amazon carries its Anthropic stake at about $239bn, on roughly $18bn invested

Two days ago this site flagged a line in Amazon's results it could not explain: "acquisitions, net of cash acquired, non-marketable investments, and other, net" of $23.986bn in the quarter against $786m a year earlier, with no breakdown in the release. Amazon filed its 10-Q on 30 October and the note explains it. The answer is larger than the question.

Anthropic. Amazon invested $8.0bn in convertible notes between the third quarter of 2023 and the fourth quarter of 2025, then $5.0bn in Series G nonvoting preferred and $5.0bn in Series H in the second quarter of 2026 — roughly $18bn of cash in total, by our addition of the filing's own figures. What it now carries:

Add the two carrying figures and Amazon's Anthropic position stands at roughly $239bn. The notes are subject to a 12% ownership cap, waivable at Amazon's election; without it they would convert to about 184 million preferred shares, and each preferred share converts into ten nonvoting common shares.

OpenAI. In the first quarter of 2026 Amazon invested $15.0bn in OpenAI Series C preferred and signed an equity commitment letter for a further $35.0bn. It funded $13.7bn in the second quarter and the remaining $21.3bn in the third — which is most of the $23.986bn line this site could not account for. The balance sheet carries $50.0bn of OpenAI Series C at 30 September 2026.

And a facility that pays out as Amazon delivers. Amazon made available to Anthropic "an aggregate facility not to exceed $20.0bn", expiring 30 months after an Anthropic liquidity event including an IPO. At inception nothing could be drawn: "as we reach certain delivery milestones of compute capacity under the amended commercial arrangement, amounts under this facility are made available for Anthropic to draw upon at its discretion." Amazon's $5.0bn Series H investment reduced it to $15.0bn.

Against that, the revenue side. AWS commitments not yet recognised stood at approximately $578bn at 30 September 2026, with a weighted-average remaining contract life of 6.1 years — including an OpenAI arrangement expanded by $100.0bn over 8 years on top of an existing $38.0bn, and an Anthropic commitment expanded by more than $100.0bn over 10 years. Both expansions, Amazon notes, include "contractual obligations related to the performance of AWS chips".

Why it matters. On 30 October this site reported Amazon's trailing-twelve-month free cash flow as an outflow of $42.4bn and asked where $24bn had gone. Here is the architecture. Amazon has put roughly $68bn of cash into its two largest AI customers, who have committed on the order of $238bn of future spending to AWS, and the revaluation of those stakes added $62.8bn to other income in nine months with $140.3bn more sitting unrealised in equity. The facility is the clearest single disclosure: money becomes available to Anthropic as Amazon hits compute-delivery milestones, and Anthropic's use for cash is buying compute.

Three cautions, because this is the kind of item that invites over-reading. These are Level 3 fair values — Amazon's own estimates against observable funding rounds, not market prices, and the filing says so. The $140.3bn in other comprehensive income is not in reported earnings. And a gain that has gone up by this much in nine months can move the other way; Amazon discloses the valuation method precisely because the number is a judgement.

Source: Amazon.com, Inc., Form 10-Q for the quarter ended 30 September 2026, filed 30 October 2026 — see "Non-Marketable Investments" and "Note 4 — Commitments and Contingencies" (SEC EDGAR).

Mistral released open weights for a trillion-parameter model, with one condition

Mistral published the open weights for Mistral Large 4, nicknamed "Le Chonk" — an improved version of the model it had put into public preview earlier in October. The post is listed on Mistral's news index under 29 October; the page itself carries no date, which is worth saying because it is the only date we have.

What Mistral states about it: roughly 1 trillion parameters, natively multimodal, a hybrid instruct-and-reasoning mixture-of-experts design, trained from scratch on European infrastructure, with three reasoning modes. Pricing is $0.70 per million input tokens, $0.07 cached, $2.10 output.

The licence is the part to read. The weights are released under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT permitting free use, modification, fine-tuning and commercial use — with one condition: a business operating a Model-as-a-Service offering needs a separate commercial licence from Mistral before using the model commercially.

On safety, two disclosures sit side by side. Mistral reports an average 96.6% refusal rate on malicious cyber requests, which it says is the highest among open models it measured, and 97.1% attack resistance on the B3 Agent Security Benchmark, up from 93.3% in the preview. It also reports that a partner saw the model refuse black-box attacks on live targets but still saw potential for real-attack uplift, and that Mistral is modifying guardrails in response. A specialised cybersecurity variant is in preview only to vetted design partners, under contractual safeguards limited to systems they own or are authorised to test.

Why it matters. "Open weights" is doing less work here than the phrase usually implies, and the carve-out is the reason: anyone may run this model, but the business of running it for other people requires Mistral's permission. That is a licence designed to let the ecosystem build while keeping the hosted-inference market — which is where the revenue is — under contract. Expect the phrase "open weights" to keep drifting in this direction.

The safety note deserves more attention than it will get. A vendor publishing that an outside partner still found uplift potential after the refusals worked, and that guardrails are being changed as a result, is disclosing an unresolved finding rather than a cleared bar. Every benchmark figure above is vendor-reported and none is independently verified; we could not open the model repository to check the licence text or the active-parameter count at source, so both rest on Mistral's own page.

Source: Mistral AI, "Le Chonk: Built for Coders, Trusted by Defenders" (Mistral Large 4 open weights), listed 29 October 2026.

A company selling AI compute filed a prospectus doubting it can continue

Exascale Labs Holdings filed a preliminary prospectus dated 30 October 2026, registering up to 12,350,000 Class A shares for resale by its SPAC sponsor and associated holders. This is not a capital raise for the company: the shares were issued at the closing of a business combination with D. Boral ARC Acquisition I Corp, completed 27 August 2026, and the proceeds of any sale go to the selling holders.

The financial disclosures, for the year ended 30 June 2026:

The business is described as asset-light: it sources GPU resources rather than owning them, and its recurring costs include "sourcing GPU resources, developing and maintaining software and platform capabilities, data center hosting, power and network services". Its own risk factors name "demand uncertainty for AI compute services" and "fluctuations in utilization rates of our GPU capacity".

Why it matters. Put this beside the first item. In the same week, Amazon carried AI-lab stakes at a quarter of a trillion dollars and Alphabet raised $147bn of external capital; at the other end of the same trade, a company selling AI compute holds $4.9m of cash — part of it in a stablecoin — against $31.9m of current liabilities and tells the SEC there is substantial doubt it survives the year. Both are documents filed days apart about the same business.

Nothing here says the company will fail; a going-concern paragraph is a disclosure requirement, not a prediction, and asset-light resellers can be refinanced. The point is narrower and worth holding onto: "AI compute" describes both a $578bn contracted backlog and a $4.9m cash balance, and the phrase itself tells you nothing about which one you are looking at. That is the most useful thing a reader can take from an ordinary week of filings.

Source: Exascale Labs Holdings Inc., preliminary prospectus (Form S-1), dated 30 October 2026 — see "Liquidity and Capital Resources — Going Concern Considerations" (SEC EDGAR).

Three items. No new filings were made over the weekend, so these come from documents of 29 and 30 October that had not been read closely: a 10-Q note, a vendor's own model page, and a preliminary prospectus. No Ark7 mention in today's items: the only place one would fit is beside a company disclosing substantial doubt about its survival, and that adjacency would imply a comparison we are not in a position to make. The advertisement on this page is unaffected and labelled, as always. How this site works.