Key Takeaways

  • Open-weight AI models grew from roughly 20% to nearly 80% market share, taking volume straight from proprietary frontier providers.
  • Token pricing is collapsing at historic speed. Altmann notes that AI tokens may be deflating faster than 2021 NFTs did during the crypto crash.
  • Developers do not need to query Nvidia Hopper or Vera Rubin clusters in an Ohio data center for everyday tasks when multi-billion parameter models run directly on local laptops.
  • Big tech hardware providers like Nvidia (with Nemotron) and Meta already distribute consumer-grade local weights that make edge execution viable.
  • Real productivity gains come from applying existing, cheaper intelligence rather than waiting for next-generation reasoning architectures.

Tokens Are Deflating Faster Than 2021 NFTs

Silicon Valley built its current financial projections around high compute prices. Venture rounds assumed proprietary lab owners would collect fat tolls on every prompt, forever. Alexander Altmann sees the opposite happening across market desks.

“You can evidently see on a whole bunch of different sources that open weight has kind of gone from a 20% market share to anything up to about an 80% market share,” Altmann observes. When model weights become commoditized downloads, the cost of generating text drops off a cliff. Altmann puts it bluntly: “It might be one of the fastest deflating assets in history. It makes those NFTs back in 2021, too.”

If tokens drop toward zero, paying high margins to proprietary APIs stops making financial sense. Startups that treat model access as a fixed premium cost will watch their unit economics get eaten by rivals running open weights for next to nothing.

You Do Not Need an Ohio Data Center

Hyperscalers are pouring billions into massive installations. They want every software engineer hooked to their infrastructure. Yet Altmann questions whether simple agent workflows actually require that heavy iron.

“And so the idea that we necessarily need to ping a bunch of Hoppers or Vera Rubens in an Ohio data center, just so that we can sort of get our agent to...” Altmann leaves the absurdity hanging. You do not need thousands of high-bandwidth GPUs across the country to parse an invoice or draft an email. Altmann points to “the ability to run, say, multi-billion parameter models locally, using whether an open weight model, or a closed weight model, but that's provided by Nvidia's Nemotron, or, you know, whatever Meta is putting out, which Meta's already put out a multi-billion parameter model that's run locally on your laptop.”

Local silicon on modern laptops already handles multi-billion parameter models without network latency, cloud bills, or privacy leaks. Shifting simple inference to consumer devices cuts server overhead and relieves bottlenecks at neocloud providers.

Stop Waiting for Frontier Reasoning Models

Founders often delay product launches while waiting for frontier labs to solve complex logic. They assume their app will only work once the next giant model drops. Altmann views this wait-and-see posture as an operational mistake.

“My big mantra is that one of the best technological changes that we experience as humans is actually not necessarily some new advanced tech,” Altmann says. “It's changing the application of existing tech.”

Running stable, smaller models inside specific enterprise workflows solves real problems today. Chasing frontier capabilities burns cash on expensive API calls while delaying user distribution.

What to Do With This

Audit your monthly API spend tomorrow morning. Identify every recurring prompt that handles deterministic tasks like classification, parsing, or data extraction, and migrate those jobs to a quantized open-weight model like Meta's Llama or Nvidia's Nemotron. If the workload runs on an employee's machine or a cheap local instance, take it off the cloud billing sheet immediately.