NPC Labs CEO: The Biggest Shift In AI Is Happening In Front Of Your Eyes (Everything You Need To Know)

Darrell, CEO of NPC Labs and core contributor to the B3 ecosystem, used the episode to bring B3 IQ out of stealth. The platform is an AI compute ownership product targeting startups, individual researchers, and investors seeking financial exposure to GPU hardware as an asset class. The founding team came out of Coinbase with a background in NFTs and infrastructure before pivoting into AI compute after observing GPU price signals as early as 2025.

The central ownership argument Darrell makes is a rent-versus-buy comparison. Renting an H200 node costs approximately 200,000 dollars per year, totaling 400,000 dollars over two years with no asset retained at the end. Purchasing the same node outright costs approximately 400,000 dollars and leaves the buyer with a physical machine. He frames the math as heavily skewed toward ownership at current market rates.

For investors without full upfront capital, NPC Labs acts as the lender using its own balance sheet, since most GPU financing in the market supports only institutional buyers like hyperscalers and large data centers. An 8x H200 node is priced at approximately 292,000 dollars on the platform. The financing model requires 30 percent down, roughly 87,000 dollars. Gross monthly earnings from renting the node to AI startups run approximately 14,000 dollars, with about 10,000 dollars per month going toward the remaining balance and a revenue share fee of approximately 4,000 dollars per month back to the platform, leaving net monthly cash to the investor of approximately 3,000 dollars. The machine reaches full investor ownership in approximately 25 months, and the stated four-year IRR on the leveraged model is 63 percent.

Supply conditions Darrell described are severe. Finding more than 500 Blackwell GPUs or a single H200 cluster of 1,000 GPUs available at one time is rare, and Blackwell cards are effectively sold out. Supply pressure has caused older generation hardware including A100s, H100s, and consumer-grade 3090s to appreciate or hold price floors, which Darrell called unprecedented in PC hardware history. H100 prices have risen approximately 40 percent since the beginning of the year, and H200 prices have risen even more. He cautioned that this appreciation is not expected to persist, as Nvidia's next generation Rubin architecture is anticipated within 12 to 18 months.

NPC Labs models a 20 percent annual decline in hardware rental rates as a conservative depreciation assumption baked into projections. The platform targets multi-quarter or multi-year direct B2B contracts to lock in rates rather than relying on volatile on-demand marketplace pricing, and sets utilization assumptions at 80 percent to reflect near full utilization during active contracts while accounting for gaps between contract periods over a four-year machine life. Under multi-year contracts the compute purchaser pays for a set number of hours regardless of actual usage, effectively guaranteeing payment. Compute contracts are currently paired on a direct one-to-one basis between investor hardware and a specific customer rather than pooled, and a pooled compute product is on the roadmap within a couple of quarters.

Hardware is hosted at a 27,000 square foot facility in Oregon, and users access purchased machines via SSH or a dashboard on bare metal single-tenant hardware. The platform covers H100s, H200s, and Blackwells. Darrell emphasized that US export controls prohibit AI startups from mainland China and other listed countries from using workloads on the platform, and the dashboard interface exists in part to satisfy know-your-customer requirements tied to those controls.

Darrell also made a broader strategic case for owning rather than renting compute. He argued that frontier AI companies like OpenAI and Anthropic subsidize API plans ahead of IPO, making them attractive short term, but that a hidden cost is proprietary data exposure since those companies do not publish model weights and do not allow local installation. Open weight models can be run on owned hardware, though he noted that publishing only weights still carries some residual risk. Current compute demand is approximately 55 percent for training with the remainder for inference, and inference demand is growing month over month as open weight models improve. Rob noted that Jensen Huang announced a partnership between Nvidia and major capital allocators including Goldman Sachs, Blackrock, Blackstone, KKR, and Apollo for AI data center financing, which Darrell cited as further evidence that compute ownership is gaining institutional mind share. A concept under ideation would structure each machine offering as an RWA vault with a minimum investment of around 1,000 dollars, abstracting away complex metrics like IRR and payback period while giving investors the option to redeem their position for actual physical machine ownership through a legal structure.