Dave Blundin's teams wanted 72 graphics processors. They had already ordered one of the systems, and someone outbid them for it anyway.
Blundin, founder and general partner of Link Ventures, used that story in a debate on the Moonshots podcast to make a point: AI computing power is now scarce. The panel disagreed about who should decide what that power is used for. Emad Mostaque, founder of Intelligent Internet, argued that governments should buy large amounts of it for universities and open science. Alexander Wissner-Gross, a computer scientist and founder of Reified, said the private sector should keep the job. The episode was recorded on October 6, 2026, and published the next day.
Mostaque's case for a "tech tree for humanity"
In AI, "compute" means the processing power, mostly racks of specialized chips in data centers, needed to train and run models. The case for public compute opened with a forecast about scale: by the end of next year, the argument went, there will be enough compute for roughly 600,000 to 700,000 runs on the scale of a Navier–Stokes-level run with a frontier model. If even 1 or 10 percent of that went to open science, why not treat it as a public good? Governments spend tens or hundreds of billions on roads, so why not tens of billions on "a tech tree for humanity"? The term comes from strategy games, where a tech tree maps which discoveries unlock which others. The idea would apply it to the questions science needs to answer.
The same turn of the discussion questioned patents and other intellectual property (IP) as a way to organize science, describing IP as an old construct from a time when people could not come together and figure out science. In its place it proposed a massive Manhattan Project for open science.
The example given was OnCo, a site built by Jude Gomila to organize the world's cancer knowledge. People are contributing to it, the discussion noted, and the expectation was that they would soon be able to donate computing power to it as well. OnCo describes itself as a public, cited and editable map of oncology. It links cancers, drug targets, treatments, pathways and trials, and it warns that entries can be incomplete or outdated. Its code is under the permissive MIT license, while its data are offered on noncommercial attribution terms, with separate licenses for commercial use.
The state does not need to run the science itself, Mostaque said. Instead it should give research institutions "stupid amounts of compute" and let researchers point it at the biggest problems. In his view, AI has been overtaking human limits in building intelligence, and science will follow. Once a problem is clearly defined, solving it becomes a question of compute. Humans still need to choose the right questions, he said, and on that he judged universities better placed than private companies.
He tied the idea to the Kyoto Vision for a golden age of science. The White House announced on October 4 that the United States and 16 other countries had endorsed it, in an initiative led by Michael Kratsios, director of the Office of Science and Technology Policy. The declaration calls for broader access to compute, data and experimental facilities and proposes new ways of funding research, including long-duration awards, rapid grants and prizes. It states shared policy goals and does not report discoveries.
Wissner-Gross: the private sector already has the compute
Wissner-Gross disagreed. He said IP is alive and well, and that "the state, this time around, did not solve superintelligence. It was solved by private labs." The state is great for certain things, he said, but he did not think state centralization was needed to cure all disease.
He raised a practical objection too. Most compute sits in the private sector, not with governments. The reply was that governments should buy compute and AI agents from private companies and hand them to researchers. Wissner-Gross was unconvinced. "It's not obvious to me what problem this is aspiring to solve," he said, since the private sector has the compute, the problems and the ability to turn solutions into products. He contrasted a statist, European-style approach with a "slightly more American" one, and noted that unlike the original Manhattan Project, this time the key resources came from private companies. If Mostaque wants a tech tree, he said, he should launch a tech tree project himself.
Wissner-Gross was most critical of rationing. If demand pushes compute prices "sky high," he said, the answer is not what he calls "thread lines," his term for government breadlines for compute. "Government rationed compute or government mediated compute is a terrible strategy," he said. Capitalism, in his view, has to feel the pain of demand outrunning supply so that it builds more chip factories and eventually reaches abundance.
The discussion then turned to a middle path: the government would buy 10 to 20 percent of compute at market rates and distribute it, while the private sector carried on. The government's stakes in Intel were cited as a precedent. The argument continued that the classic National Institutes of Health grant process does not work and that the goal should be getting as much compute as possible into American universities.
Peter Diamandis and peer review
Peter Diamandis, founder of XPRIZE and Singularity University, took a position between the two. He told Mostaque he does not think the best talent is found in academia, and that "the world's biggest problems are the world's biggest business opportunities." He also worried about the government grant process. Peer review, in which experts judge proposals, favors incremental work, he said, because a revolutionary idea can wipe out a reviewer's own standing as an expert. "Both are relevant and both are needed," he said.
Earlier in the episode, Diamandis described a case of private science funding. The longevity researcher David Sinclair had lost his NIH and National Science Foundation funding and been asked by his dean to let most of his team go. Diamandis said that after an interview on the show, he and Sinclair invited listeners to contribute through a group called Friends of Sinclair Lab, which raised $6 million, enough to replace the lost funding and more. Wissner-Gross said state funding of science is largely a World War II-era invention, and that superintelligence and recent administration policies are returning science to the older model of wealthy patrons. Diamandis compared it to the Medici family in the Renaissance.
Salim Ismail: some countries have no private option
Salim Ismail, founder of Open ExO, partly sided with Mostaque. In many countries, he said, the private sector will not supply compute. "Take Zimbabwe," he said. There it would make sense for the government to pool resources, gather energy, build data centers and make compute available to people. He also described a World Bank partnership with Google's Gemini that provides tuberculosis and diabetes screening over remote 3G cellular access, which he offered as an example of public-sector partners delivering compute directly to citizens.
Blundin's 72 GPUs
Blundin backed Ismail with what he called actual data. Entrepreneurial teams at his lab were trying to buy an NVIDIA NVL72, a rack-scale system with 72 GPUs, for something they think is very world changing. By comparison, he said, Elon Musk had just bought a million.
A year ago the system would have cost $3.5 million, Blundin said. The team had one on order for $5 million, but "somebody overbid us and scooped it, even though we had already booked the order." Their best remaining option was to lease one for $9 million over three years, and they would not own it at the end.
"The compute has all been taken by just a handful of people," Blundin said. As he described it, someone with a breakthrough idea now has two ways to get compute. One runs through the White House, via investors and dealmakers he named. The other is investment from NVIDIA's Jensen Huang or AMD's Lisa Su. Anyone outside those routes is "already frozen out," he said. That leaves universities as "one of the few ways that we might actually get some degree of meritocracy and fairness." They have their own bargaining chips, he said, in wet labs, machine shops and nanocenters, but "all the compute just ran away from them."
The scramble for compute also came up earlier in the episode. Blundin pointed to Positron, a chip company he said reached a $5 billion valuation after 16 months, with hardware that runs models without NVIDIA chips because "everything's sold out for five years." Positron's own announcement on September 10 confirmed the $5 billion valuation and $875 million in financing. It described next-generation designs built on commodity LPDDR5X memory to avoid the scarce high-bandwidth memory used in leading AI chips.
Blundin closed with a warning. If market forces are left alone, he said, an AI shopping tool called MetaMuse could become so persuasive that it "eats up every GPU that could be curing diabetes" to sell underwear or jeans to teenagers. Governments and universities, he said, are a good counterbalance and need to be a factor in deciding who gets the compute and why.
Wissner-Gross was not convinced. He asked how a conversation about abundance had come to worry that MetaMuse "of all applications" would soak up all the flops, and called it "a stupid little CUA," a computer-use agent, meaning an AI program that operates a browser or desktop the way a person would.