The Compute-Economics Inversion: Open-Weight Models and Efficiency Startups Undercut the Capex Boom
Two men sat down this week to place bets on the future of artificial intelligence, and they bet in opposite directions.
Enrique Lizaso, cofounder of a Spanish scaleup called Multiverse Computing, closed a €500M Series C on the thesis that powerful AI does not require expensive infrastructure—that the industry's central assumption is simply wrong.
Ilya Sutskever, shortly after, reportedly took a reported $5B from Nvidia to scale his compute "by an order of magnitude," on the thesis that it very much does.
Most readers assume the market will eventually pick a side. It won't. The interesting story this week is that capital is now funding both directions at once, and the squeeze is landing on the labs stuck in the middle.
Start with the anomaly. Tomasz Tunguz opened his weekend note with an image: Boris Yeltsin walking into a Houston supermarket in 1989, stunned by the ice cream aisle. OpenRouter, he argues, is that aisle for AI.
And the shopping data is strange. OpenAI's GPT-OSS 120b—an open-weight model released in August 2025—commands 36% of the volume of Anthropic's Opus 4.8, a frontier model that shipped weeks ago.
Why does a year-old open model hold a third of a fresh frontier model's traffic? Because the market never intended to converge on the best model. It intended to segment.
Last week we covered open weights as a story of US-China defiance—Chinese labs banned from Nvidia's best chips, turning scarcity into open-source dominance. This week the same trend reads differently: as enterprise unit economics.
The token market has fractured along the axes Tunguz lists—size, origin, architecture, accuracy, speed, modality. There are many flavors of AI now, and buyers are picking off the shelf rather than defaulting to the priciest jar.
The mechanism is architectural. Tunguz swapped his agent's model for Poolside's Laguna S 2.1, a 118-billion-parameter model that runs at the decode cost of a 26b model because it's mixture-of-experts: 118 billion parameters in memory, only 8 billion active per token.
His tool-call failure rate fell from 29.4% to 20.1% as active parameters climbed. Frontier-class quality is being pulled down into the local tier, on a laptop, at commodity cost.
That is the technical fact underneath the money. Now the money.
Investors paid a valuation five times its Series B valuation—€1.5B pre-money, per EU-Startups—for a company whose entire pitch is that the capex assumption is false.
The AI industry has accepted a false constraint for years—that powerful models require expensive infrastructure. That constraint is gone.
"The AI industry has accepted a false constraint for years—that powerful models require expensive infrastructure," Lizaso said. "That constraint is gone."
Multiverse's CompactifAI compresses large language models by 80–95% with what the company calls immaterial accuracy loss, using tensor networks borrowed from quantum physics. The compressed models run on smartphones, factory floors, satellites—anywhere a hyperscaler is too expensive, too slow, or not permitted.
The traction is not speculative. Multiverse reports growing annualized revenue more than 10x since its June 2025 Series B, with 96x year-over-year sales growth in Q1 2026, on a customer list that runs from Bosch and Allianz to Telefónica and the Bank of Canada.
The cap table tells you who believes this. HP Inc., Santander, Orange Ventures, BNP Paribas, Qatar Development Bank, the National Bank of Canada's venture arm—strategics, sovereigns, and banks, not just VCs chasing the frontier.
Forgepoint's Damien Henault put the thesis plainly: "The AI industry has built an extraordinary capability on top of an infrastructure model that wasn't designed to sustain it." The winners, he argues, make every GPU and every deployment dramatically more efficient.
Enterprises are voting the same way with their inference budgets. Mozilla CTO Raffi Krikorian switched to Moonshot's Kimi K3 for daily work within days of its launch—"It just seems snappier," he said, comparing it to the pricier Claude Fable.
Coinbase has reportedly said it's switching to Chinese models to cut costs. General Compute's CTO frames the shift bluntly: companies move to open weights "to keep their data private and also get better unit economics and just better AI in general that's optimized for their use cases."
Reports put some Chinese open-weight models at 60% to 90% cheaper than leading US offerings while remaining "good enough" for a large share of workloads. That phrase—good enough, cheaper, and yours—is now a fundable thesis. Multiverse's €500M is the proof.
Here's the complication, and it's the part the efficiency triumphalists get wrong: efficiency does not kill capex. It reallocates it.
The strongest version of the bearish case only holds at one layer—frontier pretraining. If an excellent open-weight base model exists, most organizations should stop trying to reproduce it, and that duplicate spending goes away.
But the other layers move the opposite way. A cheaper model gets served more often, embedded in more products, run with longer contexts, wired to more tools. Cumulative inference on a popular model can exceed the compute used to train it.
Open weights also push spending sideways—into self-hosted clusters, sovereign data centers, edge devices, and all the unglamorous complements: memory, networking, cooling, substations, power.
The evidence shows both forces at once. Open models are taking a larger share of production tokens at radically lower prices, and aggregate usage and spending are still rising. Amazon, Alphabet, Microsoft, and Meta all raised capital investment after the 2025 DeepSeek shock.
So the vulnerable thing is not "compute" in the abstract. It's the economic rent attached to a scarce proprietary model—the premium a lab can charge for an API when a comparable open-weight alternative sits one download away.
If routine inference routes to Kimi or GPT-OSS or a compressed Multiverse model, the same physical infrastructure may carry more tokens while generating less revenue per token. That is the precise pressure point: monetization per unit of demand, not demand itself.
Which brings us back to the fork. The same week efficiency became a €1.5B thesis, Sutskever's Safe Superintelligence took Nvidia's reported $5B for access to the Vera Rubin platform.
"We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so," Sutskever said. SSI has raised $7B to date at a $32B post-money valuation, and it is pursuing a "straight shot" to superintelligence with no commercial products and no near-term revenue.
Read the two bets side by side and the contradiction appears total. Multiverse says the constraint is gone; SSI is spending billions on the premise that scaling the constraint is the only thing that matters.
Both can be right, because they're serving different tokens. SSI is betting on the world's hardest problems, the frontier that still can't be commoditized. Multiverse and the open-weight tier are betting on everything else—the vast middle where "good enough" already clears the bar.
The uncomfortable position is the one in between. A lab financing an ever-larger training run to serve tokens the mid-tier can already handle is paying frontier prices to compete in a commodity aisle.
That's the compute-economics inversion. Not that the frontier disappears—it doesn't—but that its pricing power now stops at the edge of the problems only it can solve.
The frontier still serves the world's hardest tokens. It just no longer gets to serve all of them, and the tier on Tunguz's laptop keeps raising its ceiling. Competition, as he puts it, will push that forward inexorably.
The money this week understood the shape of it. One set of investors bought the ceiling; another bought the floor rising to meet it. The people who should be nervous are the ones charging frontier rent for the middle.