The report argues that when you map the money flow, Nvidia sits at the center of a circular loop: it writes equity checks and prepaid compute commitments to AI startups, which immediately recycle that capital into GPU orders. The pattern mirrors Lucent and Nortel's late-1990s vendor financing that masked collapsing end demand until the write-downs cascaded.
The editorial stresses that none of the individual deals are illegal and most appear in 10-Qs or press releases. What's new is the aggregated money-flow graph showing a meaningful slice of GPU 'demand' tracing back to Nvidia's own balance sheet — a structural risk only visible when the deals are viewed together.
The editorial singles out Nvidia's reported pre-IPO agreement to backstop unsold CoreWeave capacity as the clearest example of circularity. In that arrangement Nvidia capitalized the customer, sold it GPUs, and then guaranteed demand for the resulting capacity — closing the loop in a way that goes beyond ordinary strategic investment.
A report surfaced on Hacker News from techpowerup alleging that Nvidia has been systematically investing in AI startups — through its NVentures arm and direct strategic checks — on the implicit or explicit understanding that a large share of that capital would flow straight back to Nvidia in the form of GPU purchases. The list of recipients reads like a who's-who of the current AI boom: CoreWeave, Lambda, Applied Digital, Nebius, Wayve, Perplexity, Mistral, Figure, Cohere, and others. In aggregate, Nvidia has committed billions in equity and anchor-tenant commitments across these companies over the last two years.
The mechanics are not subtle. Nvidia writes a check — sometimes equity, sometimes a prepaid compute commitment, sometimes both. The startup books that cash and places a GPU order. Nvidia records the GPU sale as revenue, the startup records the GPUs as capex, and both sides report accelerating AI growth to the market. In the case of CoreWeave's pre-IPO bridge, Nvidia reportedly agreed to backstop unsold capacity, meaning Nvidia was effectively renting its own chips from a customer it had helped capitalize.
None of the individual transactions are illegal, and most have been disclosed in 10-Qs or press releases. What the report pulls together is the pattern: when you draw the money-flow graph with Nvidia at the center, a meaningful slice of the 'demand' driving the GPU shortage traces back to Nvidia's own balance sheet.
The comparison every finance desk is quietly making is to Lucent Technologies and Nortel Networks circa 1999. Both telecom-equipment vendors propped up their growth numbers by loaning money to the CLECs and dot-coms that bought their switches. When the buyers ran out of real customers, they defaulted, and the vendors took the write-down twice — once on the receivables, once on the demand that evaporated. Lucent went from a $258 billion market cap to a penny stock in under three years.
The counter-argument, and it is not a weak one, is that these AI startups actually ship products. Perplexity has real users. Mistral has real enterprise contracts. CoreWeave rents out capacity to Microsoft and Meta, who have genuine demand. The GPUs are not sitting in a warehouse. The question is not whether the chips are being used — they are — but whether the velocity of use justifies the price Nvidia is booking them at, or whether vendor financing is pulling demand forward from a 2027 that may not arrive on schedule.
Community reaction on the Hacker News thread was unusually sharp, with 507 points and heavy skepticism of the official narrative. One top comment compared the structure to Cisco's 2001 vendor-financing meltdown; another pointed out that Nvidia's accounts-receivable days have been quietly climbing, a leading indicator that buyers are slower to pay. A third flagged that the Microsoft–OpenAI–Nvidia triangle (Microsoft funds OpenAI, OpenAI buys Azure, Azure buys Nvidia) is a larger version of the same round-trip, just with a trillion-dollar customer in the middle instead of a Series B.
There is also a straightforward antitrust angle. If Nvidia is selectively financing customers who commit to Nvidia-only stacks, that is the kind of behavior the FTC and EU already have open inquiries about. AMD has an MI300 line that is competitive on paper for inference workloads, and Intel's Gaudi 3 benchmarks well on specific transformer shapes. If the reason you are not seeing those chips in the field is that the buyers are capitalized by Nvidia, that is a different market than the one implied by the share-price multiple.
The second-order effect is on how the rest of the industry reads the signal. When you are a cloud provider deciding whether to commit $4 billion to H200 capacity for 2026, the question 'is there a real shortage?' is load-bearing. If a nontrivial share of the backlog is Nvidia-funded demand, the shortage softens the moment Nvidia closes the spigot — and Nvidia closes the spigot the moment its margins compress. That is a reflexive loop, and reflexive loops unwind fast.
If you are a working engineer, the direct impact is on how you plan capacity and how you negotiate. Assume the GPU supply curve is softer than the public shortage narrative implies, and refuse to sign three-year reserved-instance commitments at 2024 spot prices. The hyperscalers are already quietly offering better terms on H100 capacity than they were six months ago; H200 and B200 pricing is starting to crack on the margins. If your workload can tolerate a two-quarter delay, you can probably buy the same chip for 20-30% less by mid-2026.
Second, this is the moment to actually benchmark your workload on non-Nvidia silicon. MI300X has been shipping in volume for a year. Groq and Cerebras are both real for inference. Google's TPU v5p is available to anyone who will sign a GCP commit. The CUDA moat is real for training, but for inference — which is where the bulk of production dollars will land over the next 24 months — the moat is a trench, not a canyon. Spend a sprint porting your hottest inference path to ROCm or JAX-on-TPU and see what happens. If the numbers are within 15%, your negotiating position just changed.
Third, if you run a startup that has taken Nvidia money or compute credits, read your term sheet again. Several of the deals that have leaked include most-favored-nation clauses, right-of-first-refusal on future compute, and purchase commitments that survive a down round. Those are fine when the capital cycle is up. They are a noose when it turns, because you cannot pivot to cheaper hardware without triggering a covenant.
Nvidia will almost certainly deny the framing — the company's position has been, and will continue to be, that these investments are strategic bets on the ecosystem, not round-trip revenue engineering. That may even be mostly true. But the market has started pricing in the possibility that it is not, and the AR-days creep, the widening gap between shipped-and-recognized revenue and end-customer payments, and the increasingly tangled web of equity-plus-purchase-order deals all point in the same direction. Watch Nvidia's Q4 cash-conversion ratio and the CoreWeave utilization numbers next earnings cycle. If those two metrics diverge from the revenue-growth line, the Lucent comparison stops being a bear-case meme and starts being a thesis.
This is third source reporting. WSJ does the original, Wccftech reports that WSJ reports, now Techpowerup reports what Wccftech reports about WSJ reporting.ORGINAL is the best. It tells completely different story better (not a weird twist where CEO trying to increase sales and price of their product
I thought we knew a year ago that nVidia was doing circular finance?https://www.theguardian.com/technology/2025/dec/28/nvidia-in...
Nvidia wants to ensure data center companies that are optimized for AI like Nebius stock their buildings with Nvidia chips instead of AMD or ASICs so of course it is smart business sense to partner with them or take stakes in them.Nvidia gives Nebius first access to their newest chips, ensures they
Everyone is talking about the circular financing but this is the key takeaway from the WSJ article for me.> Then he went on to describe the situation the company was in. Major cloud providers were running away from Nvidia, making plans to use alternative chips and lessen their reliance on the fir
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Reading this thread makes me realize that seemingly no one has heard the basic, mundane case for why Nvidia would want to do so much "circular financing" as a matter of totally normal business strategy. This may not be the whole story, but it's worth understanding.Because of bottlenec