Top Investor Gavin Baker’s Latest Interview: Hash Rate, Chips, and the AI Infrastructure Endgame

Top technology investor and founder of Meros Capital, Gavin Baker, recently published a lengthy interview on the "Invest Like the Best" podcast. He started with a set of numbers:

Palantir, Snowflake, and Databricks are three of the most representative star companies fostered by the American SaaS revolution over the past 10 to 12 years, each with tens of thousands of employees and having spent a full decade building ecosystems and expanding businesses.

In the past month, Anthropic has added more annualized recurring revenue (ARR) than the sum of these three companies.

"This has never happened in the history of capitalism."

Baker pointed out that Anthropic's ARR growth rate is as high as 1000%, and the only reason it cannot break through the trillion-dollar mark today is that the global supply of computing power is extremely limited. He estimated that without the bottleneck of computing power, Anthropic's actual revenue run rate could easily reach 100 to 150 billion dollars today.

In response to the tension in computing power, Anthropic has made a substantial suppression of its flagship model, Claude Opus — the number of Tokens output for the same question has been reduced by about 70%. Baker said bluntly: The number of Tokens equals the quality of the answer, and this is a commercial compromise at the expense of product experience.

OpenAI vs. Anthropic: Both are unicorns, but their spending efficiency differs by an order of magnitude

The two companies are not the same kind of business.

Anthropic's single-Token inference cost is significantly lower than that of OpenAI. When reaching a similar revenue scale, Anthropic consumes about 80% less money than OpenAI. Baker predicted that凭借极高的推理毛利率, Anthropic will achieve a strong cash flow positive this year and may have already turned positive.

OpenAI is taking a different path — aggressively locking up computing resources, and building a moat through the priority possession on the supply side. In the current context of high geopolitical uncertainty, whoever gets the computing power first has the most important strategic chips.

As for whether the valuation is reasonable, Baker's judgment is: If calculated based on "unconstrained revenue" rather than actual ARR, the current multiple of revenue is not outrageous.

Space Racks: SpaceX plans to put data centers into orbit

Baker discussed a variable in the interview that has not yet been fully priced by the market: the power shortage problem on Earth will be fundamentally solved by SpaceX's Orbital Compute around 2027-2028.

He first corrected a common misconception — "space data centers" are not floating Pentagons, but standard rack satellites that are launched into orbit one by one. A Blackwell rack weighs about 3,000 pounds, is 8 feet high, 4 feet deep, and 3 feet wide, and the rack itself is a satellite.

The technology is not mysterious: equipped with about 500 feet of solar wings on both sides, deployed in a sun-synchronous orbit to ensure continuous exposure to light for 24 hours; hundreds of feet of radiation heat dissipation plates extend behind the rack; vacuum lasers are used for data interconnection between racks, and this technology is already fully mature on Starlink satellites.

The key data Baker cited: the operating power of Starlink V3 satellites has reached 20kW, and the power consumption of ground-based Blackwell racks is about 100 to 120kW. SpaceX only needs to scale up the existing technology by a factor of 5. He said that several SpaceX engineers have clearly told him that the heat dissipation problem has been solved, and this is not a guess.

When Starship achieves full reusability and the launch cost drops significantly, all model inferences will be moved to orbit to be completed, and high-energy model training will continue to be on the ground.

Baker believes that the current high-valued ground energy infrastructure companies need to take this schedule seriously.

TSMC's production discipline is the last dam preventing bubbles

On the supply side of wafers, Baker concludes that the lifeblood of the global AI industry is one point: the group of "tough old men" in Taiwan's Science Park. They control TSMC and follow the production discipline left by Morris Chang.

Baker disclosed a detail little known to the outside world: Despite Huang Renyu's frequent flights to Taiwan to request increased production, Nvidia and TSMC have never signed any formal written contracts to this day. The huge business between the two trillion-dollar companies is based on a handshake and long-term strategic mutual trust.