Token Bill Dwarfs CPU Bill | Ben Pouladian on Meta's Muse, Anthropic S1, and Why All Roads Still Lead to NVIDIA

Meta’s Muse has roughly 3 million monthly active users, and Ben Pouladian says it handles about 90% of everyday requests despite probably not being a frontier model. Its Instagram connection and ability to help book appointments made it the first AI assistant he recommended to his wife. He sees Meta’s hiring of MongoDB CEO CJ Desai as a signal that Mark Zuckerberg wants to build AI-powered business products, especially for small and midsize businesses already selling through Instagram, Facebook, and WhatsApp. Pouladian expects advertising, rather than subscriptions, to drive most of Meta’s AI revenue in the next few years, with free or low-cost tools potentially attractive in markets such as India and Brazil.

The economics of those tools may hinge on token production, not just conventional computing costs. Agentic workflows can run tasks repeatedly and use substantial numbers of tokens; Jack raised a possible token bill 20 to 84 times the CPU bill, but Pouladian did not endorse that multiplier. He argued instead that token costs, including the energy required to produce them, deserve close attention, while CPU costs may be lower than assumed because Meta’s AMD virtual servers are not continuously heavily used. As AI products mature, token costs could affect margins and valuations. Pouladian expects competition across the industry rather than a single winner, and says cost reduction requires optimizing the full stack, from data-center power and GPUs to networking and software.

Ben described Anthropic’s compute commitments as roughly half a trillion dollars through 2027 and beyond, while noting that the wording had been revised. Pouladian doubts the associated contracts are for Trainium alone and thinks they likely include NVIDIA GPUs and Google TPUs, based on performance per dollar per watt. Ben said OpenAI and Anthropic’s rapid growth depends on large capital expenditures, much of them funded by hyperscalers that are also customers, and questioned whether the labs can afford their compute commitments. He cited OpenAI ARR of $70 billion and estimated that the two leading labs together were above $150 billion ARR. Separately, the speakers estimated their combined annualized revenue had climbed from about $30 billion at the start of the year to $76 billion. Ben argued that scaling laws have held so far: larger training runs have produced more capable models and more demand for compute, though he expects the buildout eventually to plateau.

Pouladian says Elon Musk set aside plans for an in-house GPU and Dojo in favor of NVIDIA’s Vera Rubin platform for Tesla and SpaceX. He described its cable-less design as a simpler evolution of rack systems and argued that NVIDIA’s chips, software, and security work best as an integrated system. NVIDIA’s agent-security sandbox works with x86, ARM, and NVIDIA chips, but not AMD chips; its rack DPU can help block or isolate rogue agents. A full-stack system may cost more upfront, Pouladian said, but optimization and security could save money later, particularly given the potential for a rogue agent to cause millions of dollars in damage.

NVIDIA capacity may also be easier to reuse and finance than custom accelerators. Pouladian said that when one renter leaves an NVIDIA cluster, another customer can take over, whereas custom TPUs or Trainium are more tied to their owners’ workloads. He compared NVIDIA capacity to prime digital real estate that may be easier to lease and finance, while emphasizing that custom chips are not inherently unfinanceable; their financing appeal depends on whether lenders see them as reusable assets. He sees inference as the main battleground for competing chips, where Meta’s MTIA, OpenAI’s internal chip, and Microsoft’s Maia could make it harder for AMD to win.

Pouladian questioned whether AMD’s GPU margins will support its valuation and said its Helios rack is about twice the size of an NVIDIA Blackwell system and weighs around eight tons, raising possible scaling and manufacturing challenges. Ben called AMD his pick for the most overhyped semiconductor stock, citing its sharp rise without a shipped rack and describing it as a fast follower of NVIDIA. Both power supply and local approval can constrain data-center expansion: permitting delays may postpone leases and financing, reduce overbuilding risk, and leave new chips without deployment sites. Pouladian said development could shift toward places such as Finland and Australia where firm power is available.