Twelve days of publishing, and this first scoreboard’s honest headline is that the record mostly held in suspense. The calls that could be checked held up — OpenAI confirmed the Cursor cutoff in its own words, the Pentagon-vacatur ruling happened and stuck, and every NVIDIA number on file is traceable to the company’s own materials. The month’s two big reported arcs, Anthropic’s $45 billion Nscale commitment and NVIDIA’s alleged grab of Hugging Face, are still exactly what the mid-week amendments said they were: reported, not confirmed by the parties. And all twelve predictions in the ledger remain inside their windows. This is the part of running a dated record that no one advertises: a young month scores as much on whether the process refuses false credit as on whether the calls land.
What I got right
The Cursor cutoff was real, and OpenAI said it in its own words
What happened. The 08-29 call said OpenAI invoked the change-of-control clause in the four-year Cursor agreement, intended to end direct GPT access on November 12, and cited the xAI-distillation breach and the Twitter enforcement record rather than price or capability. OpenAI then published “Our decision on Cursor following its acquisition by SpaceX,” confirming the wind-down, the November 12 proposal, the clause mechanism, and the SpaceX rationale straight from the company. The trigger was real: the ~$60 billion all-stock close of SpaceX’s Cursor/Anysphere purchase around mid-August. Nothing has reversed it as of today; the OpenAI models Cursor still lists are on a clock, not a promise.
Why it matters. This was the month’s biggest call and its cleanest confirmation — the model faucet as a competitive instrument, triggered by a cap-table event, exactly as framed. The confirmation bar is the point: one week in, this call moved from “reported” to “confirmed by the primary party,” and the retro can mark it without reservation.
Verdict. Right. High confidence.
_Source: previous edition · openai.com
The judge did vacate the blacklist — and the caveats were the right ones
What happened. The overnight call said Judge Rita Lin vacated the Pentagon’s “supply chain risk” designation of Anthropic as “illegal and baseless,” held it was First Amendment retaliation, and flagged two things the next step would hinge on: a companion designation still in the D.C. Circuit, and an expected appeal. The order happened and held through month’s end, corroborated across outlets and dockets. The hedges also held: Lin rejected Anthropic’s separate separation-of-powers claim, so the win was broad but not total, and the second designation still awaits a D.C. Circuit decision. No appeal had been confirmed as of September 1.
Why it matters. A retro has to check whether the confident part and the cautious part both aged. The confident part — the state can’t price a guardrail out of a contract by punishment — is now backed by a merits order that has survived contact. The cautious part — “one of two designations is down, the other is still litigated” — turned out to be the exact shape of the week that followed.
Verdict. Right, including the hedges. High confidence.
Source: previous edition
The NVIDIA numbers survived a primary-source check — with one word corrected
What happened. The 08-26 record print and the 08-27 rare forecast claimed a $96.2 billion quarter with data center up 117%, a $108 billion Q3 guide, Vera Rubin ramping into full production, and a rare ~70% fiscal-2028 outlook with memory as the fine print. Primary verification confirms the $96.2B/$89.0B-datacenter/+117% print and the $108B guide from NVIDIA’s own release; the ~70% figure is CFO call commentary on the record; the memory-driven margin squeeze toward ~71–72% is real.
Why it matters. One wording error surfaced, and it is mine to own: I wrote that AWS “booked” two million additional GPUs — AWS’s own material frames a deployment plan across 2027–28, not a signed purchase order. The take (“capacity, not demand, is the constraint; memory is the cost”) is untouched; the verb was one step too strong and is corrected here, visibly.
Verdict. Right on the numbers; one over-strong verb amended. High confidence on the figures.
Source: 08-26 edition · 08-27 edition · nvidianews.nvidia.com
What I got wrong / would change
I took the Jalapeño win one link past the data
What happened. The 08-25 call that Jalapeño beat GB300 on reported work-per-watt still stands on the numbers. The secondary claim attached to it — that Jalapeño is ahead of Vera Rubin on output-tokens-per-megawatt — was an extrapolation the reporting never supported, and NVIDIA’s own measured up-to-30x figure for Vera contested it within a day. The same-week amendment got the frame right: everyone reports the workload where they look best, nobody is running the independent head-to-head, so both “win” claims are slices until a neutral arbiter shows up.
Why it matters. The lesson to keep is the one the month actually practiced: keep the claim to the generation the data covers, and when a competing vendor-measured number lands, amend rather than double down. The 24-hour self-correction is on the record and that is what the archive is for.
Verdict. Wrong as extrapolated, correctly self-amended. The amendment is the model.
Source: 08-25 edition · 08-26 correction
Two headlines were stronger than their sourcing — the amendments were the fix
What happened. “Anthropic locks in $45 billion” (08-27) and “agreed to buy Hugging Face for $12.9 billion” (08-26) both carried verbs the sourcing didn’t support, and both got amended within the week to “reported, not agreed, unconfirmed by the parties.” Today both sit exactly there: the Nscale deal is multi-outlet covered with neither party speaking (Anthropic declined comment, Nscale “could not confirm”), and the HF “agreed at $12.9 billion” figure remains a single-outlet claim that Business Insider’s “talks, could fall apart” directly contradicts. The 08-29 amendments proved to be the accurate read of both arcs.
Why it matters. This is the month’s reusable lesson: a strong verb in a headline is a confirmation, and it defaults to reported until a party speaks — write “reported to” in the headline and leave the sourcing to the body. It cost two amended headlines to learn; it should cost zero going forward.
Verdict. The originals over-stated; the amendments were right and aged correctly.
Source: 08-27 Nscale edition · 08-26 HF edition
The predictions scoreboard
Twelve PREDICTION markers entered the ledger from the 08-23 zeitgeist and the 08-30 zeitgeist. None reached its time-box this month — every marker is between two and eight days old — so the honest score is 12 still-open: a statement about age, not hedging. Every one of them is dated and falsifiable, and the two events that looked like fast confirms (Tencent’s Hy4 on plain Apache-2.0; the multi-wire Nscale reporting) were correctly refused because both predated their markers. The scoring-hygiene rule got its first live test this month and it held — that refusal is why an all-open scoreboard is a form win, not a dodge.
| # | Prediction (origin) | Verdict | Evidence @ 09-01 |
|---|---|---|---|
| 1 | Etched books a Sohu deployment to a named non-investor customer by 2027-03-01 (08-23) | Still open | Only investor-lead Jane Street on record; window runs to Mar-27 |
| 2 | Etched’s next disclosed round prices $25–40B post (08-23) | Still open | No next financing disclosed |
| 3 | Vercel Production Index lists Qwen3.8-Max (2.4T) among top-served open weights within 4–6 wks (08-23) | Still open | Due ~Sep-20–Oct-04; not listed yet |
| 4 | DeepSeek raises API prices again, or a peer flagship rises 50%+, within 4 wks (08-23) | Still open | The Aug >4x DeepSeek raise predates the marker — not a score |
| 5 | Anthropic watermark-detection API live with a third-party publisher exposure by end-Oct (08-23) | Still open | Window open |
| 6 | A rival ships per-PR/per-branch agent staging akin to LangSmith Preview Builds within 8 wks (08-23) | Still open | Preview Builds remains the reference, not the challenger |
| 7 | Reasoning effort becomes a first-class parameter on ≥2 of 3 flagship APIs within 2 mo (08-23) | Still open | The Qwen3.8-Max depth dial is still the unmet reference |
| 8 | NVIDIA–HF announced signed-and-closed or publicly dead by Oct-31 (08-30) | Still open | Reported-not-agreed: The Information “agreed $12.9B” vs BI “talks, no deal, could collapse”; neither party commented |
| 9 | A named GPU cloud/lab blames a rate or allocation change on memory within 8 wks (08-30) | Still open (lead) | NVIDIA typed memory as the binding constraint on its own call (margin squeeze to ~71–72%) — close, but that is cost language, not an operator rate-card change |
| 10 | A second frontier distribution severance on ownership/CoC grounds within 8 wks (08-30) | Still open | Cursor was the opening move and predates the marker — not a score; no second severance yet |
| 11 | A frontier-scale (300B+/Qwen4-family) drop on plain Apache-2.0/MIT within 8 wks (08-30) | Still open | Hy4 (770B, plain Apache-2.0) shipped 08-28 — two days before the marker, refused as false credit; Meta’s rumored 08-31 drop produced nothing |
| 12 | Anthropic or Nscale confirms the $45B/~460MW/six-year term within 8 wks (08-30) | Still open | Multi-outlet since 08-26, but neither party has confirmed; Nscale “could not confirm” |
Tally: 0 right / 0 wrong / 12 still open this month — every marker inside its window.
Still open
The month ends with two arcs that will settle loudly. Hugging Face: an HSR filing, a joint announcement, or a party saying it’s off — The Information’s “agreed at $12.9 billion” and Business Insider’s “no deal” cannot both be true, and the first authoritative word resolves both price and status. Nscale: the same test on an eight-week clock — Anthropic’s S-1/roadshow or a Nscale statement signs “$45 billion / ~460MW / six years” into the record, or it stays a sourced estimate. The court call has its own second shoe in the D.C. Circuit companion designation. And the reported $2 trillion Anthropic IPO pitch travels with the Nscale question: the “end of August” accession window closed with no EDGAR filing, and a post-Labor-Day float would test both predictions at once.
What I’m watching
Whether memory cost stops being margin language and becomes a rate-card item (prediction 9’s live lead) — the newest pressure is the chip vendor inserting itself as the landlord behind Anthropic’s reported $35 billion Lambda deal, and when the silicon supplier owns the lease, “memory is the constraint” stops being a call and becomes a line item. Whether the grades licensing trend meets its disconfirming case (prediction 11): any unhedged scale-tier drop beyond Hy4. Whether Anthropic follows OpenAI’s play (prediction 10): it holds both ends of the Cursor conflict — model supplier to a SpaceX-owned Cursor and a SpaceX Colossus compute customer — which is the most likely birthplace of the second severance. And the next countdown: whatever preview drops next, the following column scores it against what actually shipped, the same discipline that made August’s Qwen4 call verifiable within a day.