The First Capital Signal of Nvidia’s Fracturing
General Compute didn’t just raise money. It triggered a shift.
A $400 million loan from Upper90 isn’t unusual in AI infrastructure — we’ve seen CoreWeave do it with GPUs. But this time, the collateral wasn’t a pile of H100s. It was a warehouse of SambaNova SN50 chips, purpose-built for inference, not training. That’s the first time anyone’s done it. And if you’re not paying attention to that, you’re missing the quiet collapse of Nvidia’s monopoly.
This isn’t about buying compute. It’s about capital finally recognizing that the next wave of AI isn’t built in labs with billion-parameter models. It’s running them — cheaply, at scale, and without water-cooled data centers. General Compute’s neocloud, built around SN50s, doesn’t need liquid cooling. It doesn’t need to be housed in a hyperscaler’s fortress. It can go into any rack, anywhere. And it’s 16x faster at inference than GPU clouds. That’s not a marketing claim. It’s a physics advantage.
Finn Puklowski, the CEO, put it bluntly: "This is not just, 'a cool startup got some money to buy some compute.' Like, this is the first signal of capital organizing itself and the fragmenting of Nvidia’s monopolistic dominance."
That’s the line. That’s the thesis. Upper90 didn’t bet on another GPU play. They bet on fragmentation. And they chose General Compute because the company isn’t just using alternative chips — they’re building an entire stack around them. No Nvidia lock-in. No licensing fees. No supply chain nightmares.
And they’re not alone.
TensorWave is partnering with AMD. Groq’s LPU chips are getting snapped up by startups that can’t wait for NVIDIA’s next release. Cerebras is quietly scaling. Kimi’s K3 model is outperforming OpenAI’s latest on coding benchmarks. OpenRouter and Fireworks are raising hundreds of millions to serve these models — not because they’re the biggest, but because they’re the cheapest.
The market is finally waking up to what we’ve known for years: training is expensive. Inference is where the money is. And the people who control inference — the ones who can run models without a $10 million GPU rack — are the ones who’ll win.
This loan isn’t funding hardware. It’s funding a new infrastructure paradigm.
India’s Quiet Role in the Inference Revolution
You won’t hear this in Silicon Valley, but India is becoming a critical node in this new inference economy.
HCL, India’s tech services giant, isn’t just outsourcing code anymore. They’re building inference clusters. Not for their own use. For clients who need low-latency AI at the edge — banks in Mumbai, logistics firms in Bangalore, healthcare providers in Hyderabad. They’re not buying SN50s directly. They’re integrating them into hybrid deployments with AMD and even legacy Intel Xeons, optimized for inference workloads.
Why India? Because the cost of compute there is 60% lower than in the U.S., and the talent pool is deep in distributed systems. Add to that the fact that Indian firms don’t have the same legacy dependencies on AWS or Azure. They’re building from scratch. And they’re doing it fast.
General Compute’s model — decentralized, chip-agnostic, inference-first — is the perfect fit for this market. HCL isn’t trying to compete with OpenAI. They’re trying to make AI accessible to small hospitals, regional governments, and rural schools. That’s the real scale. Not the demo at Web Summit. The real-world use case in a clinic in Pune.
This is the future: AI infrastructure that doesn’t live in San Francisco or Seattle. It lives in the cloud, yes — but also on the edge, in places where the cost of failure is measured in lives, not quarterly earnings.
Why This Deal Changes Everything
Upper90’s Billy Libby said it plainly: "The GPU market is comparatively well understood and perhaps over-bought."
That’s the quiet admission of a market that’s saturated. Every VC now knows how to finance GPUs. Everyone knows how to price them. The margins are thinning. The supply chains are locked in.
But inference? That’s still wild west. The chips are new. The architectures are evolving. The software stacks are fragmented. And the buyers? They’re not the big tech firms. They’re the startups, the regional providers, the edge operators.
General Compute didn’t just get a loan. They got a green light for a new category of infrastructure finance — one where the asset isn’t a commodity, but a catalyst.
This is the moment the AI infrastructure market stopped being about who had the most GPUs, and started being about who could run models the fastest, cheapest, and most independently.
The next billion-dollar AI company won’t be the one with the biggest training cluster.
It’ll be the one running inference on a chip no one’s heard of — in a data center in Bangalore, powered by a loan no one thought possible.
And now, thanks to General Compute and Upper90, it’s not just possible.
It’s inevitable.
