Two former executives with deep roots at Google and Nvidia have teamed up with a onetime Andreessen Horowitz partner to attack one of the artificial intelligence industry’s most stubborn problems: the chronic lack of graphics processing units.

Anjney Midha, who spent years as a general partner at the venture firm, joined forces with several ex-employees from the two tech giants to create National Compute. The new venture aims to aggregate underused AI server capacity scattered across clouds in the U.S. and allied nations, then rent it out to smaller companies and startups locked out of the hyperscaler-dominated market.

The effort comes as Nvidia itself acknowledges supply constraints will cap its growth at least through early 2028. Supply Chain Dive reported in late September that the chipmaker increased supply commitments by $160 billion quarter-over-quarter yet still faces bottlenecks spanning memory, silicon, and data-center construction. CEO Jensen Huang has described the supply chain as gigantic yet insufficient for exploding demand.

National Compute plans to tap a different source. Many enterprises lease AI servers from cloud providers but run them at far less than full capacity. The startup has already secured commitments for 750 megawatts of such slack capacity that it can resell. That scale carries enormous price tags. Similar fleets of Nvidia GPUs can command hundreds of billions in annual lease costs.

Midha told Chinese financial outlet Sina Finance on October 7 that the company holds more than $5 billion in letters of intent from potential customers. Those pre-orders signal strong appetite even before the firm finalizes its ownership structure or exact business model. The team continues gathering feedback from the market. This week it also plans to release an academic paper outlining its strategy.

The GPU crunch shows no signs of immediate relief. On the same day as the Sina report, Network World noted AMD intends to substantially boost CPU and GPU output in 2027. Yet analysts expect the broader AI compute shortage, stretching from chips to high-bandwidth memory, advanced packaging, data-center space, and power, to persist at least through the first half of next year. Premium pricing will likely remain.

Memory prices have already climbed five- to sevenfold in some cases, according to Intel executives cited in multiple reports. Micron reportedly ended production of certain lower-density GDDR7 chips, further tightening options for consumer GPUs while AI workloads consume the bulk of supply. TrendForce detailed the move in September.

Power represents another hard limit. Morgan Stanley estimates a 32-gigawatt shortfall in U.S. data-center capacity through 2028 even after accounting for workarounds. Benzinga reported the bank’s view that Nvidia and Broadcom may weather the storm better than downstream suppliers thanks to visibility into deployments and coordination with data-center builders.

National Compute’s approach sidesteps some of these constraints by focusing on utilization rather than new silicon. It pools existing but idle resources. The model resembles a broker for spare capacity, yet with commitments that give lenders and operators confidence in steady revenue. One recent X post from the io.net community highlighted exactly this dynamic: lenders prioritize long-term take-or-pay contracts over hourly usage from startups. Distributed supply from independent providers can sit outside that debt-service logic.

Plenty of competition has emerged in the neocloud and AI infrastructure space. GMI Cloud raised $263 million in equity plus a $440 million credit facility at the end of September to expand its on-demand GPU offerings globally. SiliconANGLE covered the Taiwan-based firm’s rapid revenue growth, with annual recurring revenue under contract set to jump tenfold by year-end.

Other players target different layers. CScale came out of stealth in late September with $145 million to build optical interconnects for gigawatt-scale AI clusters, backed by Nvidia and Intel Capital. Tech Startups reported the focus on keeping thousands of accelerators communicating efficiently even when parts of the network fail. Positron AI, founded by alumni of Lambda and Groq, secured $875 million in September for memory-first inference chips that promise far higher bandwidth utilization than traditional GPUs. SDxCentral detailed the $5 billion valuation.

Google and Blackstone’s joint TPU neocloud officially launched as Crux AI in September, hiring Meta’s former head of data-center engineering to scale toward 2 gigawatts of capacity. Data Center Dynamics quoted new chief development officer Alan Duong on the need for an end-to-end chain that holds from site selection through decades of operation.

Yet many of these ventures still chase the same scarce Nvidia GPUs or build alternatives that take years to reach volume. National Compute bets that better matching of supply and demand can deliver capacity faster and cheaper for the companies currently shut out. Startups in particular face brutal economics. Some must raise fresh capital simply to lock in multi-year compute reservations that exceed their current bank balances.

The original briefing on the company’s formation appeared in The Information. It captured the frustration rippling through the industry as hyperscalers tighten control over GPU allocations. Smaller players pay premiums or wait months for access. National Compute wants to loosen that grip without waiting for new fabs or power plants to come online.

Midha also founded another firm this year called AMP, likewise focused on broadening access to AI servers. The parallel efforts suggest a concentrated push to expand the effective supply of compute. Whether National Compute can convert its 750 megawatts of commitments and $5 billion pipeline into actual delivered capacity remains the test. The team has not yet settled ownership splits among the Google, Nvidia, and Apple alumni involved.

Industry watchers note the timing feels urgent. AMD’s promised 2027 ramp offers hope, but packaging, memory, and electricity constraints point to continued tightness. Nvidia’s own guidance shows revenue growth of 70 percent for its fiscal 2028 yet falls short of what customers want. In that gap sits opportunity for brokers, aggregators, and efficiency plays.

So far the market has rewarded bold infrastructure bets. Reflection AI, another startup with former Google DeepMind talent, raised billions including $800 million from Nvidia and recently unveiled an efficient open-weight model. PC Mag reported on October 6 that its Beam model delivers strong performance at a fraction of the compute cost of Chinese rivals. Demand for intelligence keeps climbing faster than hardware can scale.

National Compute won’t manufacture new chips. It won’t break ground on gigawatt data centers. Its contribution may prove more prosaic yet equally necessary: wringing higher utilization from resources already purchased and deployed. In an industry obsessed with the next accelerator generation, the former executives are choosing to optimize what exists today.

That choice reflects a maturing view of the bottleneck. Compute scarcity remains the binding constraint. Talent from the companies that created the shortage now flows toward fixing it. Success for National Compute would mean more startups can train models, more researchers can run experiments, and the AI boom can broaden beyond the handful of organizations with privileged access. The next few quarters will show whether pooling idle servers delivers the relief so many are seeking.