Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

Provider Policy & Trust

This Week in DeAI: The Proof Problem

Nvidia lost the man who made open weights policy, a forensics report went viral unverified, and this week every big claim arrived without proof attached.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

An empty leather chair at the head of a conference table in a darkened corporate boardroom, a single closed laptop in front of the empty seat, photographed from the doorway — Nvidia's CEO seat vacated, the week's lead story. Illustration: DeAI
An empty leather chair at the head of a conference table in a darkened corporate boardroom, a single closed laptop in front of the empty seat, photographed from the doorway — Nvidia's CEO seat vacated, the week's lead story. Illustration: DeAI

November 7–13 was the week the proof ran out ahead of the claims. The biggest story of the cycle — Jensen Huang's resignation from Nvidia — arrived as a two-paragraph corporate release and left every consequence unmeasured: no successor, no stated mandate on the $12.93 billion Hugging Face acquisition, no answer on who now carries Huang's export-control advocacy. The loudest story — a viral report claiming six Chinese open-weights models carry Blackwell initialization fingerprints — arrived with 26,000 likes on the announcement and no code in the repo. And the week's quietest policy shift — xAI turning watermarking on by default for the Grok API — was the rare story that shipped with its evidence attached: documentation, a response header, and an audit-logged opt-out, all checkable the day it landed.

That spread is the week's theme. Every layer of the open-weights and private-inference stack spent this week making claims, and they did not arrive with equal evidentiary weight. A forensics report asked to be believed on statistical elegance alone. A fine-tune asked to be checked — and gave you everything needed to check it. A provider asked you to trust an unverifiable claim about quality unaffected by watermarking. A price index watched rows move with no announcement from anyone. The pattern this publication exists to cover — the gap between claimed and verified — did not narrow this week. It got more expensive to ignore.

Last week's hub ended on the discipline of verification: Moonshot's weights deadline checked instead of assumed, the H200 hosting claim recomputed instead of repeated. This week stress-tested the same discipline from the other direction — what happens when the claims outrun the checking apparatus, and when the person whose judgment anchored the entire hardware layer leaves the building.

Key facts

The resignation: when the strategy and the strategist part ways

Start with the resignation, because everything downstream of it inherits the uncertainty. Nvidia's board announced on November 12 that Huang resigned as President, CEO, and board member effective immediately, ending a 33-year run, and that CFO Colette Kress — CFO since September 2013 — would serve as interim CEO while a permanent search considers internal and external candidates. Harvey C. Jones, a venture investor and former Synopsys chief executive, became the company's first-ever board chair in the same announcement. Q3 earnings stay scheduled for November 17, which now doubles as the first public test of the interim tenure (our full coverage).

For this beat, the load-bearing fact is not the personnel change itself — it is how much of the open-weights ecosystem's hardware foundation was Huang's personal project. The company he leaves is mid-flight on commitments that matter to anyone who runs open models: the $12.93 billion Hugging Face acquisition announced September 3, under which Huang publicly promised that Nvidia compute would not be required to build on or deploy through the platform; the Nemotron open-weights line, which put models like Nemotron 3.5 Transcribe 30B into weights-available distribution; and the reported hundreds of billions in Blackwell-class bookings that fund the datacenters open-weights inference is served from. None of these has changed. All of them now belong to an interim CEO without a stated mandate on any of them.

The second open question is export control, and it connects directly to the week's loudest claim. Huang was Washington's most prominent corporate voice against tightening chip restrictions, having co-signed a public letter arguing that open weights strengthen American AI leadership. That advocacy did not exist independently of his position. Whoever inherits the permanent seat inherits the lobbying role, and Washington's appetite for tightening Blackwell export rules may read the landscape differently with him out of the corner office. The DePue report — whatever its evidentiary status — is a public argument that the controls are already leaking. A weakened advocate for the status quo plus a viral leakage claim is how policy windows open.

Hugging Face chief executive Clem Delangue publicly committed to continuing the open-platform fight in the days after the resignation, and the deal has not been rescinded. But a $12.9 billion acquisition mid-integration depends on board-level priorities, and the person who set those priorities is gone. Watch the November 17 earnings call for how the open-weights and Hugging Face commitments are described when Huang is not the one describing them.

The forensics report: a claim built to outrun its evidence

The week's most viral story is also its best case study in what verification is not. Will DePue's Unemployment Labs report, published November 12, claims that recovered PyTorch initialization seeds and streaming-multiprocessor layouts match six Chinese open-weights models — DeepSeek-V4, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ling-3.1 variants — to Blackwell-class GPUs, implying the use of export-controlled hardware despite US restrictions. The announcement drew 26,000+ likes and 3 million+ views within the beat window. Velocity, not verification (our PULSE coverage).

The technical method deserves to be understood separately from the claims built on it, because if it holds up, it is a real forensic technique. PyTorch initializes weights with a deterministic generator whose output depends on a seed and on how many GPU threads share the work — and the thread count is a function of the GPU's streaming multiprocessor count. A Blackwell B200/B300 with 148 SMs produces a 303,104-thread layout; an H100 with 132 SMs produces 270,336. Weights that training never touches, like embedding rows for tokens the model never saw, preserve their initialization values exactly. The report says it recovered seeds and matched thread layouts against those preserved values, with correlations above 0.999 on untouched weights.

What would verify it: an independent rerun of the seed search against the published checkpoints, a public repository with the extraction code, a named on-record reviewer, or responses from the four named labs. None of those exist as of this writing. The report publishes no code, describes a single analysis pass not yet re-run, and says its own evidence "supports, without proving" the training inference — a further step removed from the initialization finding, since initialization and training can happen in different places. Whether any listed entity acquired Blackwell-class GPUs in violation of US rules is a question for regulators with subpoena power, not for a statistical match in an embedding table.

The export-control angle is where this story meets the resignation. If checkpoint forensics ever becomes reproducible, weight provenance becomes a real due-diligence tool — and a political instrument, since a provable fingerprint method turns "the controls leak" from talking point into testable claim, at exactly the moment Nvidia's loudest advocate for the current regime left. Until replication exists, though, every specific in the report is a claim, and this publication grades it exactly that way: the tracker pipeline will not touch it, and neither should your procurement decisions.

The watermark: the rare claim that shipped with its evidence

Against that backdrop, xAI's watermarking change is instructive precisely because it is checkable. As of November 7, text output from seven current Grok model variants — grok-4.7 down through grok-4.3 and the grok-4.20 family — carries an invisible statistical watermark on every xAI API endpoint, per xAI's own developer documentation. Integrators can read a per-request x-watermark header. The opt-out belongs to team admins, applies per model, propagates within five minutes, is audit-logged, and requires confirming a legal exception. Detection is statistical and gated: xAI scores text and reports watermarked, not watermarked, or uncertain, with detection API access granted by application (our full coverage).

xAI frames the change as EU AI Act compliance, twenty-five days before the Article 50(2) synthetic-content marking deadline on December 2 — the first confirmed default-on watermark policy at a major frontier API, and the first hard answer to how providers will actually meet that deadline. The week-ahead brief flagged the December 2 runway three days before xAI moved; expect the compliance-announcement cluster it predicted to continue.

One claim in the docs stays unverified: xAI's assertion that watermarking does not change response quality or speed. That is the company's own assurance, not an independently measured property — the same claimed-and-unverified shape as everything else this week, just wearing better documentation.

The fine-tune: verification as a product feature

Arete-27B is the week's counterexample — a claim structured to be checkable. Nous Research released the model under Apache 2.0 on Hugging Face the evening of November 6: a supervised fine-tune of Alibaba's Qwen3.8-27B, trained on about 52.8 million tokens with its Inkling model as teacher and Grok 4.7 as judge, aligned — in the lab's own description — to Western values. On Nous's own 803-prompt Propaganda_Eval, answers containing Chinese state framing fell from 22% to 0.9% across five runs, and the fact score rose from 2.32 to 2.98 out of 5, with self-reported capability checks showing no meaningful regression (our full coverage).

Every one of those numbers is a vendor number on a vendor rubric. What separates Arete from the DePue report is not the strength of the claim — it is the packaging. The weights are live as BF16 safetensors, with 38 community quantizations published within hours. The eval set is public. The full training recipe — tokens, epochs, steps, teacher, judge — is disclosed on the model card. The same community that cannot re-run a seed search without code can re-run this eval today, with a non-Nous judge, and the 22%-to-0.9% claim either survives that or it does not.

The structural takeaway is bigger than one model: the value layer in a strong open model is now removable with a small, documented fine-tune, and the same move works in any direction anyone wants. A fine-tune that scrubs one jurisdiction's framing can install another's. The property worth tracking is not this particular alignment target but the evidence regime: open weights plus public evals plus disclosed recipes is what a checkable claim looks like, and the week's other stories show how rare that still is.

The incidents: trust with availability attached, and metered

Two stories rounded out the week's trust theme from different directions.

NEAR AI reported unauthorized access to its internal systems on November 9 and paused NEAR AI Cloud inference as a precaution, with its status page showing Chat API at 92.59% uptime for the week. The company says customer workloads stayed inside hardware-isolated enclaves and that NEAR Protocol was not affected — but it has not published a postmortem or stated what systems the intruder touched, and Crypto Briefing's report that the access was "addressed and closed" is the outlet's account of NEAR AI's statements, not an independent audit. The lesson generalizes past one provider: hardware isolation protects the prompt path, not the operator plane. A confidential-inference provider can be running exactly the TEE architecture our private-inference explainers describe and still go dark for a week while it audits its own front doors. Availability belongs in the private-inference decision alongside attestation — and so does the provider's incident record.

OpenAI's December change is quieter but more broadly felt: Plus, Pro, and Business plans move to a usage allowance measured in tokens and shared across ChatGPT and Codex, per OpenAI's help center, while six GPT-5 and o3 snapshots leave the API on December 11. Coding-agent traffic now draws down the same pool as chat, so budgeting a subscription in sessions rather than tokens no longer works. The same month Novita AI retires multiple hosted models on November 26 on about two weeks' notice — the recurring fall pattern where the endpoint you depend on is a lease, not a possession. Pin what you use, mirror the weights you cannot afford to lose, and keep the migration guide current.

The rest of the week, briefly

  • Inkling 2 Small and Ripple-65B Preview bookended the licensing question in opposite directions: Thinking Machines claimed the top western open-weight score at 594B parameters with no repo on Hugging Face, while Ricursive Intelligence shipped Apache 2.0 weights now at a claimed $32M total R&D cost. The model you can legally download still beats the model with better claims when the decision is what to run this week. As of Friday, the Inkling weights promised for "next week" remain unshipped — the K3.1 Pro deadline check set the standard for how to score that promise.
  • The Monday Price Index delta found 8 of 17 OpenRouter rows moved with no shared direction, and the week's largest move broke the basket's longest-stable row: llama-4-maverick jumped +87% on input after three flat weeks. DeepSeek V4 Flash 0731 printed a fourth distinct shape in four observations. No single-observation move on a volatile row has yet predicted the next Monday — price against this week, not last week, and check the cheapest-LLM reference before quoting anything.
  • Tracker Sweep Cycle 7 appended 19 entries across 13 providers, surfaced candidates from a record 18 projects, and logged the funnel's first positive evidence: two pricing-fetch corroborations against chutes.ai and venice.ai, each hashed. The grade census still reads zero verified across 17 providers. The gap between claimed and verified is not closing on its own; it is the space this publication's harness exists to shrink.
  • Together AI's serverless card listed 25 open-weight chat models as of November 10, from GLM-5.3 Flash at $0.15/$0.50 to Kimi K3 at $3.00/$15.00, with cache discounts of 80–98% on the Flash tiers and Batch mode at half price — the durable reference behind this week's moving rows.
  • Yvligo, loglog.ai's DeepSeek V4.1 Flash fine-tune that runs X ad campaigns as a tool-using agent, kept the specialist-versus-frontier debate running on vendor self-evals with public traces — the beat's healthiest version of an unverified claim.
  • Qwen's "The becoming begins" teaser pulled 1.4k likes in hours and the community reads a Qwen 4 announcement. The teaser is verified; no repo exists yet. Same shape as the rumor cluster the week-ahead brief has carried for three weeks: undated, unconfirmed, and correctly unhedged-against.

What the week changed for builders

  • Treat every viral claim as a verification problem with a checklist. Code or no code? Re-run or self-assessment? Named reviewers or anonymous release? The DePue report fails all three; Arete-27B passes all three. The tracker harness operationalizes exactly this triage, and its zero-verified census is the honest scoreboard.
  • Hardware-layer risk is now a single point of failure. Nvidia supplies effectively all GPU capacity open-weights inference runs on, and its open-weights commitments were one person's. Nothing has changed — yet. Watch the November 17 earnings call and any statement about the Hugging Face acquisition before assuming continuity.
  • Compliance defaults are arriving from the API side first. xAI's default-on watermarking, OpenAI's metered shared pool, and the 10% data-residency premium on regional endpoints confirmed in Friday's brief all push the same direction: the closed-API path is getting measured, priced, and marked. The open-weights path gets whatever governance you build — including none. Read open-weight versus open-source and what are open-weight models before assuming "open" settles anything.
  • Check the incident record, not just the architecture. NEAR AI ran TEEs and still paused entirely. Attestation protects the prompt path; it does not audit the operator plane. The TEE deep-dive and zero-retention guide cover the technical layer — availability and postmortem discipline are separate diligence items.
  • Leases expire. Six API snapshots gone December 11, one shared token pool for ChatGPT and Codex, Novita models retired on two weeks' notice. If your workload depends on someone else's endpoint, the self-hosting math deserves an annual re-run — this quarter, maybe more often.

The week in numbers

  • 33 years / 0 days — the run Jensen Huang's resignation ended, and the notice period the board gave (NVIDIA Newsroom).
  • $12.93B — Nvidia's pending Hugging Face acquisition, announced under Huang in September, now owned by an interim CEO (our coverage).
  • 6 models / 0 repos — the checkpoint-forensics report's named targets, and the number of verification artifacts it published (the report, our PULSE coverage).
  • 26,000+ likes / 3M+ views — engagement on the forensics announcement in the beat window — velocity, not verification.
  • 7 variants / 25 days — Grok models watermarked by default, and the runway left to the EU AI Act Article 50(2) deadline when xAI moved (xAI docs).
  • 22% → 0.9% — state-framed answers on Nous's 803-prompt eval after the Arete-27B fine-tune, with weights, eval, and recipe public (our coverage).
  • +87% — llama-4-maverick's input-price jump on the Monday delta, breaking the basket's longest-stable row after three flat weeks (the delta).
  • 8 of 17 — OpenRouter Price Index rows that moved on November 9, with no shared direction (the delta).
  • 19 entries / 0 verified — tracker sweep cycle 7's writes, and the grade census it left unchanged (the sweep).
  • 92.59% — NEAR AI Cloud Chat API uptime for the week of the unauthorized-access pause (status page, our coverage).
  • 1 pool / 6 snapshots — OpenAI's December shared token allowance across ChatGPT and Codex, and the GPT-5/o3 models leaving the API December 11 (our coverage).

The spokes: what each story established

What to watch

Three things, in order of consequence. First, the November 17 Nvidia earnings call — the first test of the Kress interim tenure, and the first public reading of how the Hugging Face acquisition and Nemotron commitments are described when Huang is not the one describing them. Second, whether anyone replicates or refutes the DePue seed-search method: an independent rerun against the published checkpoints either mints a real forensic tool and hands export-control hawks a testable leak claim, or it folds into the week's pile of unverified viral claims. Third, the December 2 EU AI Act Article 50(2) deadline, now nineteen days out — xAI moved first with default-on watermarking, and the compliance-announcement cluster the week-ahead brief predicted should start filling the calendar. Keep an eye, too, on the Inkling 2 Small weights promise now past its "next week" window, and on whether any independent lab re-runs the Arete-27B eval with a non-Nous judge — the cheapest verification win available on this beat right now.

Questions

Did Jensen Huang resign from Nvidia and who replaces him?
Yes. Nvidia's board announced on November 12, 2026 that Huang resigned as President, CEO, and board member effective immediately, ending a 33-year run. CFO Colette Kress is interim CEO while a permanent search runs; Harvey C. Jones became the company's first board chair. Q3 earnings remain scheduled for November 17.
Is the Blackwell checkpoint forensics report verified?
No. Will DePue's Unemployment Labs report claims PyTorch init seeds and streaming-multiprocessor layouts match six Chinese open-weights models to Blackwell-class GPUs, but it publishes no code, describes a single analysis pass not yet re-run, and says its own evidence 'supports, without proving' the training claim. It remains a claim.
Does the Grok API watermark text by default?
Yes — this one is verified. Since November 7, 2026, xAI's developer documentation states seven Grok model variants return statistically watermarked text on every endpoint, with a team-admin-only per-model opt-out. xAI frames it as EU AI Act compliance ahead of the Article 50(2) deadline on December 2.
What is Arete-27B and are its claims verified?
Arete-27B is Nous Research's Apache-2.0 fine-tune of Qwen3.8-27B. On Nous's own 803-prompt eval, state-framed answers fell from 22% to 0.9%. Those are vendor numbers on a vendor rubric — but the weights, eval, and full recipe are public, so independent reruns are possible, which is what separates it from the week's unverifiable claims.

Sources

  1. NVIDIA Announces CEO Transition — NVIDIA Newsroom
  2. Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs — Unemployment Labs
  3. Output Watermarking (xAI developer documentation) — xAI
  4. NousResearch/Arete-27B — model card, weights, and eval — Hugging Face
  5. NEAR AI Status — major outage in progress, Chat API 92.59% uptime — NEAR AI (status page)
  6. Understanding ChatGPT and Codex usage limits — OpenAI Help Center
  7. OpenRouter API — list models — OpenRouter

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →