Compute Race - Bryant McGill

# Compute Race **Entity class:** Strategic competition **Domain:** Artificial intelligence / science / political economy **Maturity:** Developed ## Definition **Compute Race** The compute race is the competition to accumulate chips, data-center capacity, model-training infrastructure, and the energy required to operate machine intelligence at scale. ## Mechanism and Consequence Because capability and deployment depend on material infrastructure, public arguments for deceleration must be evaluated alongside the actual compute purchases and construction programs of the institutions making them. ## Relationships - [[wiki/Compute Sovereignty|Compute Sovereignty]] - [[wiki/AI Infrastructure|AI Infrastructure]] - [[wiki/Energy-Compute Nexus|Energy-Compute Nexus]] - [[wiki/AI Superpower|AI Superpower]] ## Source Dossiers - [[research/AI Is Now So Good They Will Pay You to Quit|AI Is Now So Good They Will Pay You to Quit]] - [[research/Super Intelligence Week Index of Statements from the White House Luncheon|Super Intelligence Week Index of Statements from the White House Luncheon]] - [[research/Jensen Pushes Back on Doomers Moonshots Live|Jensen Pushes Back on Doomers Moonshots Live]] ## Simple Reminders, Quotations, and Thoughts > "Nobody is building more compute today than the people asking to be slowed down. All of the other narratives deflect blame and make it sound like AI is so powerful: ‘I have no idea how to fix it. It’s not my fault. It’s just because the technology is so powerful.’" > **— Jensen Huang**, *MOONSHOTS Live, September 25, 2026* [[reminders/AI-Tech Doomers/The Builders Asking to Slow AI Are Still Building the Most Compute by Jensen Huang|The Builders Asking to Slow AI Are Still Building the Most Compute by Jensen Huang]] > "[With a formal proof] you can click a button and it just checks it formally and says right or wrong. So it’s highly unlikely to be wrong. Humans can’t do that because it’s so freaking exhaustive." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/Formal Proof Can Make Exhaustive Mathematics Machine-Checkable by Emad Mostaque|Formal Proof Can Make Exhaustive Mathematics Machine-Checkable by Emad Mostaque]] > "There’s a theorem called Fermat’s Last Theorem. The official proof of that is 127 pages. Anthropic formalized that proof. It’s 13 million lines of code." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/Formalizing Fermat Required Thirteen Million Lines of Code by Emad Mostaque|Formalizing Fermat Required Thirteen Million Lines of Code by Emad Mostaque]] > As of the last month and a half, the models have gone beyond their training data definitively. Before Navier–Stokes, OpenAI released ten proofs of famous math problems, and at least two of them — [one] conjecture being the main one — [are] definitively novel, elegant, and never seen before, the way they did that proof. [AI solved] Navier–Stokes in a bit of a more mechanical way, but again in an incredibly novel way. [The Navier–Stokes proof] is quite difficult to read because it’s like AI slop Claudism when you’re reading it. It took three days for me to even read the paper, and I haven’t gone through the formal stuff. I’ve talked to enough of my mathematician buddies to say: yeah, a human could have done [the AI-generated Navier–Stokes proof] given twenty or thirty years. > **— Adapted from Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/AI Has Begun Producing Novel Mathematics Beyond Its Training Data by Emad Mostaque|AI Has Begun Producing Novel Mathematics Beyond Its Training Data by Emad Mostaque]] > "Starting August 29th, OpenAI had a new type of model training. They had an existing model that they figured out a trick with that made it go from 10% on FrontierMath, a benchmark, to 60% overnight, and [it] just started solving these math problems one after the other." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/One Training Trick Raised FrontierMath Performance Sixfold Overnight by Emad Mostaque|One Training Trick Raised FrontierMath Performance Sixfold Overnight by Emad Mostaque]] > "We’ve reached this point where [the latest AI] models have suddenly gained another capability, which is the capability to conceptualize. There’s something internal going on within them." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/AI Models Have Begun to Conceptualize by Emad Mostaque|AI Models Have Begun to Conceptualize by Emad Mostaque]] > "You can say, “I want a boat made of sugar cubes and a boat made of concrete and a boat made of wood,” and [a video model] will actually accurately show them dropping into the water: the sugar cube will kind of float, and then it will melt. It’s learning these things. It’s learning representations." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/Video Models Can Simulate How Different Materials Behave in Water by Emad Mostaque|Video Models Can Simulate How Different Materials Behave in Water by Emad Mostaque]] > "[Math is] one of the most formally verifiable [domains], but even outside of that we’re seeing human-level performance drop. Two days ago, the AIs tied the best humans in superforecasting. There are some humans who are better than others at forecasting things like outcomes of political campaigns [and] who’s going to win elections. Every year [Metaculus has] a superforecasting championship where the best human teams come together and forecast stuff. AI has just tied the best human, and obviously it will beat the best human now. [Forecasting] is something that’s open-ended, because math is kind of deterministic. You’d think a computer should be able to do math better than a human. This high-level conceptual math is a bit different, but [AI has] got there, and now it’s got there in open-ended domains." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/AI Has Tied the Best Human Superforecasters by Emad Mostaque|AI Has Tied the Best Human Superforecasters by Emad Mostaque]] > "The actual solution [to Navier–Stokes] isn’t that interesting in real-world terms, but what other people have built around that, like physics-based neural networks, will be incredibly useful for building new materials." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/Navier-Stokes Tools Can Help Build New Materials by Emad Mostaque|Navier-Stokes Tools Can Help Build New Materials by Emad Mostaque]] > [Solving great technical problems often produces valuable secondary inventions:] we get to the Moon, and then we have pens that can write upside down. > **— Adapted from Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/Moonshots Create Valuable Inventions Around the Main Goal by Emad Mostaque|Moonshots Create Valuable Inventions Around the Main Goal by Emad Mostaque]] > "The potential for these new discoveries is huge. If you’re a pure mathematician, then math would be solved. I think a lot of physics problems will be solved, and that will be great for humanity. The tools that come out will again be great for humanity." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/Solving Mathematics and Physics Will Be Great for Humanity by Emad Mostaque|Solving Mathematics and Physics Will Be Great for Humanity by Emad Mostaque]] > Across mathematics, forecasting, media generation, software, agentic work, and robotics, AI is reaching a new level of capability at the same time. > **— Adapted from Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/AI Is Reaching New Capability Across Six Domains at Once by Emad Mostaque|AI Is Reaching New Capability Across Six Domains at Once by Emad Mostaque]] > I understand what is happening in AI. We reached a level of optimization, and the internals of the models reached a scale at which progress became predictable. Now several exponential trends are overlapping and making the combined rate of progress even more exponential. > **— Adapted from Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/Scientific Acceleration/Overlapping Exponential Trends Are Accelerating AI Progress by Emad Mostaque|Overlapping Exponential Trends Are Accelerating AI Progress by Emad Mostaque]] > "The costs are dropping, but then the next generation of chips are coming out. Vera Rubin, Nvidia’s new chip, is a ten-times improvement in cost for inference." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/AI Infrastructure/New Chips Can Improve Inference Cost Tenfold by Emad Mostaque|New Chips Can Improve Inference Cost Tenfold by Emad Mostaque]] > "NVIDIA accelerated computing delivers value across generations. CoreWeave's NVIDIA V100 GPUs are still running customer workloads nearly a decade after Volta launched, even as CoreWeave brings Vera Rubin NVL72 into production. That's the strength of the NVIDIA platform: infrastructure that keeps earning for years, and the flexibility to put the right GPU on the right workload." > **— Ian Buck**, *From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI, September 30, 2026* [[reminders/AI Infrastructure/AI Infrastructure Can Keep Earning Across Hardware Generations by Ian Buck|AI Infrastructure Can Keep Earning Across Hardware Generations by Ian Buck]] > "Agentic coding is a complex workload: long contexts, high concurrency and token volumes where cost per token decides what we can ship. Having all of it on one platform, with NVIDIA and CoreWeave engineers who work the hard problems alongside ours, matters more to us than any single spec." > **— Silas Alberti**, *From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI, September 30, 2026* [[reminders/AI Infrastructure/Cost per Token Decides What Agentic AI Can Ship by Silas Alberti|Cost per Token Decides What Agentic AI Can Ship by Silas Alberti]] > "CoreWeave's deployment of Vera puts 128 CPUs and 11,264 cores in a single rack, enough for more than 11,000 concurrent environments at one core each." > **— NVIDIA**, *From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI, September 30, 2026* [[reminders/AI Infrastructure/One AI Rack Can Run More Than Eleven Thousand Concurrent Agent Environments by NVIDIA|One AI Rack Can Run More Than Eleven Thousand Concurrent Agent Environments by NVIDIA]]  ## ASI and RSI Timeline Research Routes This node is connected to the developed research ontology for [[research/ASI and RSI Timeline Research Moonshots|ASI and RSI Timeline Research Moonshots]]. The route distinguishes the source transcript from canonical wiki objects and lets later Reminders inherit an already connected conceptual graph. - **Master ontology:** [[ASI and RSI Timeline Ontology]] - **Model Architecture and Compute Economics:** [[ASI and RSI Timeline Ontology#Model Architecture and Compute Economics|complete route]] · [[World Knowledge and Reasoning Kernel Separation]] · [[Recursive Architectural Self-Improvement]] · [[Model Self-Redesign]] · [[nanochat]] · [[NanoGPT Speedrun]] · [[Algorithmic Improvement Without Data Scaling]] · [[Compact Reasoning Kernel]] · [[Externalized Model Knowledge]] - **Physical Intelligence and Scientific Infrastructure:** [[ASI and RSI Timeline Ontology#Physical Intelligence and Scientific Infrastructure|complete route]] · [[Project Suncatcher]] · [[Physical Superintelligence]] · [[Geoengineering]] · [[Weather Engineering]] · [[ASML]] · [[One-Nanometer Lithography]] · [[Two-Nanometer Lithography]] · [[Computational Substrate]]