Zendoric
← Back to the day · September 3, 2026

The GPU access problem NVIDIA can't solve alone

🕒 Published on Zendoric: September 3, 2026 · 10:20

✨ AI-generated · how it's made

The article, by Gaurav Sharma (CEO of io.net, a decentralized physical infrastructure network, or DePIN, and previously CTO of the same company, as well as having held infrastructure roles at Amazon, Binance and Agoda), analyzes NVIDIA's recent program allowing AI startups to hand over a percentage of their…

The article, written by Gaurav Sharma (CEO of io.net, a decentralized physical infrastructure network, or DePIN, and previously the same company's CTO, having also held infrastructure roles at Amazon, Binance and Agoda), examines NVIDIA's recent program allowing AI startups to pledge a percentage of their future revenue in exchange for access to GPU capacity they could not otherwise afford. The program launched in July and was received as a step toward the 'democratization' of AI infrastructure, but according to the author it reveals precisely the opposite: that the very companies that depend on NVIDIA's chips cannot get hold of them. In fact, weeks after its launch, NVIDIA reportedly paused parts of the program for fear it would attract antitrust scrutiny, which the author reads as a sign that even the company itself is aware of the concentration of power it would entail.

The backdrop the article lays out is the scale of capacity hoarding: at GTC 2026, NVIDIA disclosed $1 trillion in locked-in orders for its Blackwell and Vera Rubin chips, with capacity reserved through 2027. That capacity, the author notes, has been secured mostly by hyperscalers and frontier AI labs years in advance, leaving startups, researchers and well-funded teams competing for the leftovers, usually at inflated prices. At the same time, a considerable share of GPU capacity sits idle in data centers that lack efficient mechanisms to aggregate or resell it.

The core of the argument plays on the double meaning of the word 'fix': NVIDIA has 'fixed' the market in the sense of rigging it, through massive supply deals that gave priority to hyperscalers, and now intends to 'fix' it in the sense of repairing it — while charging for the solution. The author describes the program's mechanism as a chain in which NVIDIA finances cloud partners in exchange for a share of the revenue that hardware generates, those partners resell the capacity to startups, and NVIDIA ends up collecting both on the chip sale and on a recurring slice of the resulting revenue, with the middleman still in place and no change in who decides where compute capacity goes.

To put the problem in context, the article draws historical parallels: the deregulation of telecommunications in the 1990s, which allowed virtual network operators to resell capacity from several carriers without owning any infrastructure of their own; and the arrival of cloud computing, which took servers out of basements and turned them into a resource anyone could rent by the hour, making companies such as Airbnb, Slack, Spotify and Instagram possible. The author warns, however, that the cloud has also drifted toward concentration: AWS, Azure and Google Cloud today control around two-thirds of global cloud infrastructure, and the allocation of GPUs within that market depends more on business relationships and contract size than on open bidding.

As a solution, the author proposes creating a kind of 'clearing house' for compute: a layer of companies sitting between the big cloud providers and the rest of the market, whose job would be to locate underused GPUs — in both centralized data centers and decentralized networks — aggregate them and make them available without requiring multi-year commitments or a prior relationship with a hyperscaler. He cites as precedents Stripe, which made banking infrastructure accessible to anyone able to write a few lines of code, and virtual network operators in telecoms.

The article closes by pointing to what, in his view, is at stake: companies that never get founded because their founders have to build their roadmap around the compute they can afford rather than the best technical bet, research teams that fit their experiments to the compute hours available instead of to what the science demands, and ambitious ideas that go unfunded because everyone knows in advance that infrastructure costs will choke the company. For the author, NVIDIA's acknowledgment of the access problem is a positive sign, but the real answer is not for a single company to open a door and charge a toll every time someone walks through it — it is a market with enough doors that nobody has to wait for permission to build.

The article is worth reading with the caveat that it is an opinion column by an executive whose own company, io.net, operates precisely as a decentralized GPU compute network — that is, the kind of 'access layer' the piece proposes as the solution to the problem it describes. That does not invalidate his arguments about the concentration of capacity at NVIDIA and the hyperscalers — a dynamic widely discussed in the industry — but the proposed solution should be taken as coming from an interested party rather than as neutral analysis.

🔗 Related on Zendoric

Sources & references