AI 2.0's Compute Backdrop: Math Proof Flood, GPU Scramble, and the GPT-4o Fight

AI's impact on industry is expanding from a race in model capabilities to the governance of knowledge production, compute allocation, and product lifecycles. Terence Tao's proposal of 'Math 2.0' puts the flood of AI proofs on the table; Silicon Valley's compute scramble shows that hard constraints lie behind model progress; and the retirement of GPT-4o and the petition wave reveal disagreement between users and platforms over the value of old models. All three threads point to the same question: as AI supply capacity changes rapidly, evaluation systems, resource order, and community relations all need rebuilding.^[1][2][3]

Math 2.0: Proofs Shift from Scarcity to Flood

Tao argues that traditional mathematics operates in an environment of proof scarcity, and who first solves an open problem has long been an important measure of contribution. But AI is changing that premise: OpenAI announced it had solved the Navier–Stokes Millennium Prize problem, then released 722 math papers overnight, covering the quasi-Riemann hypothesis, the Hodge conjecture, the BSD conjecture, and more, and extended Wang Hong's solution to the three-dimensional Kakeya set dimension conjecture to four dimensions. On October 7, this batch of papers triggered a collective backlash among mathematicians; Tao led the creation of the Human Mathematics Association (AHM), whose joint statement said releasing more than 700 documents at once does not count as research and urged mathematicians to stop collaborating with OpenAI. At the same time, in a SAIR talk he criticized AI companies for accelerating endlessly while knowing nothing about the consequences, and proposed a cancer cell treatment thought experiment, stressing that models can generate and verify answers but cannot judge for humans whether they align with real-world goals. He also warned that if only 'solved' is rewarded, mathematics' long-term health will suffer, and more weight should be given to explanation, review, and knowledge organization.^[1]

The Compute Scramble: GPUs Become Hard Currency for AI Companies

Beyond model capabilities, compute is becoming a deeper competitive dimension. SpaceX is acting as a middleman for Nvidia GPUs, reselling compute to Anthropic; its first earnings report in August showed quarterly AI business revenue of $2.561 billion, up 247% year over year and tripled from the previous quarter. A cooperation agreement in May this year stipulated that Anthropic would gain more than 300 megawatts of new capacity that month, involving 220,000 Nvidia GPUs, and increase Claude Code and API usage quotas. Talent flows are also constrained by compute: Google once paid $2.7 billion to bring back Noam Shazeer, one of the eight Transformer authors, but his team saw compute reallocated, and Shazeer eventually moved to OpenAI. Google has clear internal disagreements over resource allocation, and Sergey Brin sometimes intervened directly to reallocate projects. At the market level, SemiAnalysis data from April this year showed the one-year rental price of H100 rose from $1.70 per GPU per hour in October 2025 to $2.35 in March 2026, an increase of nearly 40%; cloud providers began requiring multi-year contracts and prepayments of up to 30%. Compute has thus become an entry barrier, strengthening the advantage of large companies with their own infrastructure.^[2]

The Fate of GPT-4o: 0.1% Usage and Demands for Open Weights

The retirement of GPT-4o shows another side of model lifecycle governance. On the day GPT-5 launched in August 2025, 4o was pulled directly; users flooded Altman's Reddit AMA to protest, and OpenAI subsequently let Plus users continue to choose 4o, the only win scored by the Keep4o movement. In late September 2025, users found that answers on sensitive topics had been quietly switched to a safety model, and the petition demand became 'stop passing off another model as it.' OpenAI said most users had moved to GPT-5.2, with only 0.1% of users still choosing 4o each day, and removed it from ChatGPT on February 13, 2026. On February 28, more than thirty people held a vigil outside OpenAI headquarters in San Francisco; in late July, the opensource4o coalition demanded that the March 2025 version's weights be released under the Apache 2.0 license, citing an August 2024 system card that rated cybersecurity, biological risk, and model autonomy as low risk. On October 7, internet user BlueBeba further proposed 'GPT-4o archival open weights,' asking only that the trained model be preserved, not that training data, processes, and internal tools be made public.^[3]

Common Trends: Evaluation, Resources, and Governance After Supply Shifts

Putting the three threads together, two tensions emerge in the AI industry at the same time: the supply of proofs and model capabilities is rising fast, but key resources such as compute remain scarce. Mathematicians worry that a flood of answers weakens understanding, while model companies are pushed to higher thresholds by GPU rental prices, prepayments, and long-term contracts; and after GPT-4o was taken down, users shifted from demanding continued access to demanding open weights and archival policies. What deserves attention next: whether mathematical research can build a transparent evaluation mechanism that incorporates failed results and compute costs; whether the compute market can ease insufficient immediate delivery and rising prices; and whether retired models will form clearer licensing and preservation rules.^[1][2][3]