Nvidia does not report cloud revenue, operate data centers under its own brand, or bill enterprises by the compute hour. Clark Tang, a partner at the venture firm Altimeter, argues in a guest essay published on Tae Kim’s Substack that the chipmaker has nonetheless assembled every structural piece of a hyperscaler except a hyperscaler’s balance sheet, and is now building that piece too.
Tang’s definition of a hyperscaler is functional rather than brand-based. It does two jobs: it evens out how customers pay for infrastructure, turning hardware spend from a lump sum into a rented, recurring cost, and it hides the underlying primitives behind software so customers consume capacity instead of managing machines. Amazon, Microsoft and Google built that model around CPUs, storage and multi-tenant virtualization, earning operating margins Tang puts in the mid-30s to low-40s percent by pooling scale across many customers at once.
Training and inference broke that formula, in Tang’s account. A synchronous training run stalls entirely if one node lags, and inference economics run on tokens per watt rather than how many virtual machines fit in a rack. Neither maps onto the redundant, general-purpose fleets built for the cloud era. Nvidia filled the gap on two fronts at once. On infrastructure, it built DSX OS and Mission Control for fleet operations, DSX reference designs and Omniverse digital twins for site design, and Dynamo for inference serving. On capital, Nvidia and financing partners including Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR (through vehicles such as a Brookfield fund and KKR’s Helix) share revenue and credit support so utilization gets pooled across operators much as multi-tenant cloud software once pooled it across customers. Tang counts six such independent financing platforms and roughly $500 billion in third-party capital committed to the arrangement.
The mechanism is concrete: financing structured to only qualify Nvidia hardware makes that hardware the preferred collateral for infrastructure investors chasing unlevered yields, which Tang puts in the 7 to 8 percent range. That is the moat. Capital concentrates on Nvidia not out of loyalty but because the underwriting has been built specifically around that hardware. Reframing the resulting lock-in as a “moat dressed up as a risk,” in Tang’s phrase, is the strongest part of his case: an ecosystem where independent lenders will only fund Nvidia-based sites at this scale is evidence the platform functions as designed, not proof it is fragile.
The weaker link is collateral durability. Tang cites CoreWeave repricing six-year-old A100 chips upward by 25 percent in July, and contracting them through 2029, as evidence the hardware ages like an aircraft rather than a smartphone. One vendor’s pricing move in a single month is not a depreciation curve, and the essay does not address what happens to that collateral value if a future architecture makes older Nvidia silicon meaningfully worse on tokens per watt, the exact metric Tang says now sets the economics.
The tension the essay leaves open is that Amazon, Google, Microsoft and Meta are each funding custom silicon, respectively Trainium, TPUs, Maia and MTIA, specifically to cut their dependence on Nvidia’s margins over time. If any of those chips reaches parity on the workloads that matter, the compute Tang describes as “fungible and transferable” stops being fungible, and the financing platforms built around Nvidia’s reference designs inherit that risk directly.
For operators weighing multi-year compute contracts, the practical takeaway is that financing terms tied exclusively to Nvidia hardware are underwriting a bet that Nvidia’s software moat outlasts the hyperscalers’ own silicon roadmaps. That bet gets retested with every new generation of custom chips those hyperscalers ship.
Clark Tang, a partner at Altimeter, writing in a guest essay published on Tae Kim’s Substack on August 12, 2026.
