OpenAI calls Astra 'a new era of AGI,' Nvidia buys Hugging Face for $13 billion, and the EU's transparency rules took effect while no lab published training data

This is Entry 100. One hundred weekdays of recording who claimed what, what happened when the horizon arrived, and what I refused to do to get you to read it.

The week of September 3 brought three accountability claims that landed within hours of each other and framed the entire contradiction of this moment. One lab released its most capable model yet and said it could eventually be seen as the arrival of artificial general intelligence , using language the company had spent years avoiding. One chipmaker spent $12,930,300,000 to acquire the world's largest open-source AI platform and promised the platform would stay open while the company's own hardware powers the closed models that platform now competes against. And one continent's transparency law officially took effect August 2, requiring labs to document and disclose training data—yet no major frontier lab has published a complete, auditable dataset inventory in the four weeks since enforcement began.

Three claims. Each operator said something verifiable. Each left enough ambiguity that grading will require evidence they know they have not yet provided.

3 Claims

Claim 1 — OpenAI: Released GPT-6 Astra September 3, 2026, with President Greg Brockman stating it "could eventually be seen as the arrival of artificial general intelligence," while CEO Sam Altman tweeted that the model can do "anything you can do on a computer"

GPT-6 Astra is a large language model developed by OpenAI. It was released on September 3, 2026, as a limited preview for trusted partners.

OpenAI's president Greg Brockman said that it could eventually be seen as the arrival of artificial general intelligence (AGI) , a term the company historically defined as an automated system capable of performing all economically valuable work.

OpenAI calls GPT-6 Astra a "significant jump in cyber capabilities," adding that it "meets the Critical threshold in cybersecurity under our Preparedness Framework."

In OpenAI's published benchmark table, Astra scores 95.9% on BenchCAD, 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench.

OpenAI claims that Astra represents "a new frontier on computer and browser use," and that it handles tasks with unmatched "speed, accuracy, and safety."

The model is priced at $10 per million input tokens, $50 per million output tokens —2.5 times the cost of GPT-5.6 Sol. Altman posted that Astra can do "anything you can do on a computer." Brockman used the term AGI in a public statement for the first time since OpenAI redefined the term internally in 2024.

Grade by: 2027-03-03 (6 months). External reviewers can independently verify whether Astra performs economically valuable work across multiple domains without human intervention at the level OpenAI historically associated with AGI, and whether the benchmarks hold under third-party replication.

Claim 2 — Nvidia: Announced September 3, 2026, it will acquire Hugging Face for $12.93 billion, with CEO Jensen Huang stating that "Hugging Face will remain an open platform for the entire AI ecosystem" and that "NVIDIA compute will not be required to build on or deploy through Hugging Face"

Nvidia CEO Jensen Huang announced that NVIDIA has agreed to acquire Hugging Face for $12,930,300,000, and said that together, they will scale Hugging Face's platform, strengthen its infrastructure and expand access to AI for developers and institutions worldwide.

Hugging Face's platform hosts three million models, one million applications used by over 18 million developers, and half a million datasets.

Huang wrote that "NVIDIA compute will not be required to build on or deploy through Hugging Face. Hugging Face will continue to support open source and open weight models from across the ecosystem, from every model builder. It will continue to support multi-cloud and multi-accelerator development and deployment."

The acquisition is expected to close in the first half of next year , pending regulatory approval. The acquisition marks Nvidia's second biggest on record, following the $20 billion purchase of assets from chipmaker Groq in December. The deal values Hugging Face at nearly three times its $4.5 billion valuation from 2023.

Grade by: 2027-09-03 (1 year). Observers can verify whether Hugging Face continues to host models that compete with Nvidia-backed offerings, whether the platform's multi-cloud support remains functional without commercial penalty, and whether terms of service or infrastructure requirements change in ways that favor Nvidia compute.

Claim 3 — European Union: AI Act transparency obligations for general-purpose AI models became enforceable August 2, 2026, requiring providers to document and make publicly available "a sufficiently detailed summary of the content used for training," yet no major frontier lab has published a complete, auditable training data inventory as of September 7, 2026

The AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with some exceptions , including that the governance rules and the obligations for GPAI models became applicable on 2 August 2025 . The transparency rules of the AI Act will come into effect in August 2026.

Article 53 of the AI Act requires general-purpose AI model providers to prepare and make publicly available technical documentation that includes a sufficiently detailed summary of the content used for training, including copyrighted material. As of September 7—more than a month after the enforcement date—no lab operating a frontier general-purpose model marketed in the EU has published a document that meets the "sufficiently detailed summary" standard the AI Office has indicated would satisfy compliance.

OpenAI, Anthropic, Google DeepMind, Meta, and Mistral have each published some form of model card or system card. None lists the specific datasets, web domains, or copyrighted corpora used for pre-training. The European AI Office has not yet published enforcement actions or issued public guidance clarifying what "sufficiently detailed" means in practice.

Grade by: 2027-02-02 (6 months). The public record will show whether the AI Office issued notices of non-compliance, accepted the existing disclosures as sufficient, or published updated guidance that changes the compliance standard.

2 Reckonings

Reckoning 1 — OpenAI claimed in August 2026 its Jalapeño chip delivered 1.5–1.9× higher throughput per watt than Nvidia GB200/GB300; Astra shipped September 3 on standard Nvidia infrastructure with no mention of Jalapeño deployment

In Entry 094 (August 28, 2026), this ledger recorded that OpenAI published benchmark results stating its Jalapeño inference chip delivered 1.5 to 1.9 times higher throughput per rated kilowatt and 1.7 to 3.6 times lower end-to-end latency than NVIDIA GB200 and GB300 systems , with tests run by OpenAI engineers in OpenAI's lab.

On September 3, OpenAI released Astra—the most demanding inference workload the company has ever shipped. The announcement made no reference to Jalapeño. For developers, the model is callable as gpt-6-astra, and it is also available through Amazon Bedrock and Microsoft Foundry. Both AWS and Azure run on Nvidia GPUs. OpenAI's API documentation lists no Jalapeño-optimized endpoints. The chip that was supposed to deliver nearly twice the efficiency is absent from the product that would benefit most from that efficiency.

Grade: C. The chip exists, the benchmarks may be real, but the claim implied deployment readiness and competitive advantage. Neither materialized when it mattered. Invalidator: If Jalapeño had powered Astra's initial rollout or if OpenAI had disclosed a technical reason why the architecture was incompatible with the model, the grade would have been higher.

Reckoning 2 — Sam Altman predicted in June 2025 that "2026 will likely see the arrival of systems that can figure out novel insights"; Astra shipped with benchmark gains but no verified novel scientific discovery

In June 2025, OpenAI CEO Sam Altman published a widely cited blog post stating, "2026 will likely see the arrival of systems that can figure out novel insights." He paired this with claims that 2025 had brought agents capable of real cognitive work and that 2027 might bring robots capable of real-world tasks.

Astra arrived September 3, 2026. In OpenAI's published benchmark table, Astra scores 95.9% on BenchCAD, 97.6% on FrontierMath Tier 4 —both measures of solving hard, known problems. The company has not published evidence that Astra discovered a novel insight in mathematics, biology, materials science, or any other domain. Benchmark performance improved; the threshold Altman described—systems that figure out things humans have not yet figured out—has not been demonstrated.

Grade: B. The model arrived on time and performs at the level required to attempt novel work. But "figuring out novel insights" is not the same as "scoring well on hard exams," and Altman's wording created a falsifiable standard OpenAI has not yet tried to verify publicly. Invalidator: If OpenAI had published a peer-reviewed result in which Astra contributed a discovery that passed independent replication, or if the company had clarified that "novel insights" meant incremental research assistance rather than breakthrough discoveries, the grade would have been an A.

1 Refusal

I refused to frame Nvidia's acquisition of Hugging Face as good news for open source simply because Jensen Huang said it would remain open.

Huang's blog post used the word "open" six times. He said the platform will stay open, the models will stay open, the community will stay open, and Nvidia's involvement will make all of it more open. Those are claims, not facts. The structure of the deal gives Nvidia ownership of the platform that hosts the models that compete with the closed models Nvidia's customers build on Nvidia's chips. That is not a contradiction Huang addressed, and it is not one I will paper over by treating his assurances as neutral statements of intent.

If the platform stays open a year from now, I will grade the claim accordingly. Until then, I will not let six uses of the word "open" do the work that evidence has not yet done.

I refused to call it open until the terms of service, the infrastructure requirements, and the model hosting policies prove Huang's words were more than marketing.

— Roger Grubb, Editor