# Compute Scarcity **Entity class:** Infrastructure constraint **Domain:** Artificial intelligence / science / political economy **Maturity:** Developed ## Definition **Compute Scarcity** is the condition in which demand for chips, memory, inference, and training capacity exceeds available supply even as algorithms and hardware become more efficient. ## Mechanism and Consequence The scarcity moves the AI bottleneck down the stack. Model access gives way to chip allocation, energy delivery, cooling, construction, and capital as the limiting factors of deployment. ## Relationships - [[wiki/AI Infrastructure|AI Infrastructure]] - [[wiki/Energy-Compute Nexus|Energy-Compute Nexus]] - [[wiki/Compute Sovereignty|Compute Sovereignty]] - [[wiki/AI Price-Performance|AI Price-Performance]] ## 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 > "Energy is already the bottleneck. The United States is short roughly sixty gigawatts of energy for the next two years." > **— Peter Diamandis**, *MOONSHOTS Live, September 25, 2026* [[reminders/AI Infrastructure/Energy Is Already the AI Bottleneck by Peter Diamandis|Energy Is Already the AI Bottleneck by Peter Diamandis]] > "Both things can be true: a generalized model can drive a car or inhabit and build a robot, while a specialized model may still perform the task more efficiently. The missing question is how much compute the general model used. Compute will remain starved; chip prices have risen, and high-bandwidth memory has increased fivefold in price. AI is so valuable that people want the compute, so using more compute for the same task is a serious failure." > **— Dave Blundin**, *MOONSHOTS Live, September 25, 2026* [[reminders/AI Infrastructure/General Models Must Still Justify Their Compute Costs by Dave Blundin|General Models Must Still Justify Their Compute Costs by Dave Blundin]] > "The AI bottleneck is shifting from models down the stack to compute, and then eventually to energy." > **— Salim Ismail**, *MOONSHOTS Live, September 25, 2026* [[reminders/AI Infrastructure/The AI Bottleneck Moves From Models to Compute to Energy by Salim Ismail|The AI Bottleneck Moves From Models to Compute to Energy by Salim Ismail]] > The prediction that agentic AI will transform the economy within two years depends upon the cost falling dramatically. At today’s prices, deploying enough high-end agents to produce that transformation would be too expensive. > **— Adapted from Tom Bilyeu**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/AI Access/Agentic Economic Transformation Depends on Falling Costs by Tom Bilyeu|Agentic Economic Transformation Depends on Falling Costs by Tom Bilyeu]] > "We took the cheapest DeepSeek model and made it beat OpenAI’s last frontier model using this harness, and the cost is 50 times lower." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/AI Access/A Cheap Model Beat a Frontier Model at One-Fiftieth the Cost by Emad Mostaque|A Cheap Model Beat a Frontier Model at One-Fiftieth the Cost by Emad Mostaque]] > "You can build a website as good as Fable with this incredibly cheap model if you architect and harness it correctly — for 100 times cheaper." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/AI Access/AI Architecture Can Cut Software Costs One Hundredfold by Emad Mostaque|AI Architecture Can Cut Software Costs One Hundredfold by Emad Mostaque]] > "With the additional chips, with the harnesses and others, you should expect a 100-times decrease [in price] for an equivalent performance between now and one year from now." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/AI Access/Equivalent AI Performance Could Become One Hundred Times Cheaper in a Year by Emad Mostaque|Equivalent AI Performance Could Become One Hundred Times Cheaper in a Year 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]] > "While you need frontier models to be close to capability, you don’t want [to etch them]. The moment you have a good-enough model, then you do want that, because it costs a fraction of the price and it’s 100 times faster, a hundred times cheaper. When you use a model that goes at 1,000 words a second versus 50, it’s night and day. [AI output] is instantaneous, and you’re like: well, I’m the slow one. The latest MiniMax H3 model is eight seconds now to generate a 15-second clip. As a human, I need 15 seconds to see this, and then I’ll need more time." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/AI Access/Good-Enough Models Can Run One Hundred Times Faster and Cheaper by Emad Mostaque|Good-Enough Models Can Run One Hundred Times Faster and Cheaper by Emad Mostaque]] > "Five years from now will be the biggest infrastructure boom we’ve ever seen in our lives, both from government programs to stimulate it — because it’s the easy thing for them to stimulate — [and] because it’s kind of needed. The infrastructure is a bit creaking." > **— Emad Mostaque**, *Impact Theory with Tom Bilyeu, 2026* [[reminders/AI Infrastructure/AI and Robotics Could Produce the Largest Infrastructure Boom of Our Lives by Emad Mostaque|AI and Robotics Could Produce the Largest Infrastructure Boom of Our Lives by Emad Mostaque]] > "Claude Opus 5.5 delivers performance equivalent to Fable 5.1 at literally half the price." > **— Peter Diamandis**, *MOONSHOTS Live, September 25, 2026* [[reminders/AI Access/Frontier AI Performance Fell to Half the Price by Peter Diamandis|Frontier AI Performance Fell to Half the Price by Peter Diamandis]] > "On our private suite of 2,441 finance tasks covering Q&A, extraction, analysis, and forecasting, Claude Sonnet 5.5 scored ahead of Sonnet 5 and used about 121k tokens per answer, where Sonnet 5 used 497k. On our analyst search and retrieval work, it was better than Sonnet 5 in almost every way." > **— Joe Poirier**, *Introducing Claude Sonnet 5.5, September 28, 2026* [[reminders/AI Access/AI Cut Financial-Analysis Token Use by More Than Seventy-Five Percent by Joe Poirier|AI Cut Financial-Analysis Token Use by More Than Seventy-Five Percent by Joe Poirier]] > "Sonnet 5.5 rechecks data in source documents, catching errors that Sonnet 5 failed to spot. Compared to the last model, Sonnet 5.5 was more accurate, 2.4x faster, and used 12% fewer total tokens." > **— Yashodha Bhavnani**, *Introducing Claude Sonnet 5.5, September 28, 2026* [[reminders/AI Access/AI Became More Accurate, Two-Point-Four Times Faster, and More Efficient at Once by Yashodha Bhavnani|AI Became More Accurate, Two-Point-Four Times Faster, and More Efficient at Once by Yashodha Bhavnani]] > "With millions of Rovo-assisted actions powering our customers' workflows each month, execution speed is critical. Claude Sonnet 5.5 will allow teams to run their Rovo Agents up to 30% faster than they could with Sonnet 5." > **— Jamil Valliani**, *Introducing Claude Sonnet 5.5, September 28, 2026* [[reminders/Agentic AI/Enterprise AI Agents Can Now Run Thirty Percent Faster by Jamil Valliani|Enterprise AI Agents Can Now Run Thirty Percent Faster by Jamil Valliani]] > "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]] > A frontier NVIDIA GPU system was on order for three million dollars and due in December. Before delivery, another buyer offered five million dollars, and the system was sold for two million dollars more. That is what frontier-compute scarcity looks like in a secondary market. > **— Adapted from Dave Blundin**, *MOONSHOTS Live, October 2026* [[reminders/AI Infrastructure/A Three-Million-Dollar GPU System Was Resold for Five Million by Dave Blundin|A Three-Million-Dollar GPU System Was Resold for Five Million by Dave Blundin]] > Recursive self-improvement may become the highest-value use of a frontier laboratory’s own tokens. If stronger models generate more future value than outside applications, socially valuable work can lose compute to the development of the next model generation. > **— Adapted from Alexander Wissner-Gross**, *Moonshots, October 7, 2026* [[reminders/Machine Succession/Recursive Self-Improvement May Become the Most Valuable Use of Frontier Tokens by Alexander Wissner-Gross|Recursive Self-Improvement May Become the Most Valuable Use of Frontier Tokens by Alexander Wissner-Gross]] > Executives are still learning how to absorb today’s models while laboratories prepare much stronger ones and the compute market tightens behind them. Organizational adoption can lag twice: first behind capability, and then behind the infrastructure required to catch up. > **— Adapted from Emad Mostaque**, *Moonshots, October 7, 2026* [[reminders/Economic Transition/Executives Are Adapting to Models the Frontier Has Already Left Behind by Emad Mostaque|Executives Are Adapting to Models the Frontier Has Already Left Behind by Emad Mostaque]] > When frontier compute is scarce enough to be overbid, leased at extraordinary cost, or allocated through privileged relationships, scientific merit no longer determines who can run an experiment. Access politics becomes part of the research method. > **— Adapted from Dave Blundin**, *Moonshots, October 7, 2026* [[reminders/AI Infrastructure/Compute Scarcity Turns Scientific Merit Into Access Politics by Dave Blundin|Compute Scarcity Turns Scientific Merit Into Access Politics by Dave Blundin]] > AI deployment can become constrained by memory rather than models. When startups become valuable because they route around high-bandwidth-memory bottlenecks, RAM architecture is no longer a supporting detail; it is part of the capability frontier. > **— Adapted from Dave Blundin**, *Moonshots, October 7, 2026* [[reminders/AI Infrastructure/RAM Can Become the Binding Constraint on AI Deployment by Dave Blundin|RAM Can Become the Binding Constraint on AI Deployment by Dave Blundin]] > Token efficiency and cost efficiency are different metrics. A model can emit fewer visible tokens while performing more internal computation for each one, so fewer tokens do not necessarily mean a cheaper or more efficient system. > **— Adapted from Alexander Wissner-Gross**, *Moonshots, October 7, 2026* [[reminders/AI Infrastructure/Token Efficiency and Cost Efficiency Are Different Metrics by Alexander Wissner-Gross|Token Efficiency and Cost Efficiency Are Different Metrics by Alexander Wissner-Gross]] > Compute scarcity can delay mundane automation even when the software exists. If disease research, scientific simulation, robotaxis, customer service, and domestic robots compete for the same chips, capacity flows first toward the applications expected to create the most value per token. > **— Adapted from Alexander Wissner-Gross**, *Moonshots, October 7, 2026* [[reminders/AI Infrastructure/Compute Scarcity Can Delay Mundane Automation for Higher-Value Research by Alexander Wissner-Gross|Compute Scarcity Can Delay Mundane Automation for Higher-Value Research by Alexander Wissner-Gross]] > Recursive self-improvement is constrained by more than compute. Memory, energy, evaluation, and implementation still slow each iteration; the singular transition arrives when AI compresses the innovation cycle faster than institutions and people can follow it. > **— Adapted from Salim Ismail**, *Moonshots, October 7, 2026* [[reminders/Machine Succession/Recursive Self-Improvement Is Constrained by More Than Compute by Salim Ismail|Recursive Self-Improvement Is Constrained by More Than Compute by Salim Ismail]]