Anthropic (FlashProve), Google and AMD ($13.4B) — verification, forced control and the compute bill
Anthropic publishes FlashProve, an open-source zkSNARK prototype that proves a response came from committed weights. Google turns AI tools (Gemini, Notebook, Classroom and Search) off by default on K-12 accounts and warns that schools that fully disable AI lose Search. AMD closes Q3 with $13.4 billion in revenue (+45% YoY) and Data Center at $8.5 billion (+97%).
On the same day someone offers a cryptographic proof of which model answered, the most-used search engine forces schools to accept AI or go without search, and the chip maker feeding both reports the largest quarter in its history.
🔒 Anthropic (FlashProve): 110 tokens/s of proof on one H100
On 2-3 November Anthropic’s Frontier Red Team released FlashProve on GitHub (Apache 2.0). It is a research prototype of provable inference: the provider publishes a commitment to the weights and then generates a proof that the response came from that specific model without revealing the weights. They tested it on Llama 4 Maverick (400B MoE, ~17B active per token, 4-bit quantized). On a single H100 it reaches about 110 tokens per second of prefill in batches of 24, with ~18.8-19 MB proofs that a CPU verifies in about 17 seconds (24 threads) or 21 seconds (one thread).
The overhead is the weak point: proving costs 70x to 95x the GPU time of normal vLLM on prefill (and ~20x on decode in some configurations). The weight commitment takes ~39 minutes and covers about 16 million requests. Soundness, completeness and zero-knowledge properties are formalized in Lean. It is not intended for production and does not run on Claude; it is a research artifact.
🛡️ Google: no AI, no Search in schools
On 3 November Google announced that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be turned off by default on K-12 student accounts. Administrators will have to explicitly enable them. Rollout is expected in early 2027. The condition is clear: Search and AI are “deeply integrated”. If a school fully disables AI features, students lose access to Google Search and see a message from their school. Chris Phillips, VP of Education, confirmed this to the New York Times.
Google is also bringing Guided Learning to Search AI Overviews so students do not receive the direct answer. Several districts (New York, Los Angeles) had already restricted or banned generative tools. The reaction was immediate: politicians and educators accused Google of using its Search monopoly to force acceptance of AI. Google later clarified that services can be enabled separately, but Search will still carry integrated AI.
📊 AMD ($13.4B): Data Center nearly doubles
On 3 November AMD reported fiscal Q3 2026: $13.4 billion in revenue (+45% YoY), the largest quarter in its history. The Data Center segment contributed $8.5 billion (+97% YoY and +27% sequential), driven by EPYC and the launch of Instinct MI450. It accounts for approximately 63-64% of total revenue. Non-GAAP EPS $2.09. For Q4 it guides about $17 billion (±3%), which would imply ~66% YoY growth. Lisa Su said she expects Data Center to exceed 80% of company revenue in 2027.
Among the customers of the first Helios rack-scale solutions are Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud. Data Center growth covers almost the entire revenue increase of the quarter. Client & Gaming fell 8% and Gaming 40%. The machine serving both cryptographic proofs and AI-powered searches is billing at record pace.
The connecting thread
A tool is published that, in theory, lets you cryptographically verify which model generated a response, but at a compute cost 70-95 times higher than serving it. At the same time the dominant search engine tells schools they cannot keep Search without accepting integrated AI. And the chip supplier feeding both the labs and the hyperscalers reports that its Data Center business has nearly doubled in a year and already accounts for nearly two-thirds of its revenue. Trust is sold as optional; the compute that makes it possible is not.