The investor test for AI has changed: robots must earn their keep, and moats must be real

Sell the work, not the machine

Mayfield's advice to Physical AI founders is blunt: "1. Sell labor, not robots. Get paid for the work itself." The point is that a robot is not the product; the hour of work it does is. Mayfield puts the prize in those terms too, saying it believes Physical AI will usher in the era of labour-as-a-service across a global labour market worth more than $30 trillion. That framing changes what a founder has to prove. It is no longer enough to show a machine that moves well in a demo; the company has to show the work it can sell, hour after hour, at a price a customer will pay. m:…9315

The pricing is unsettled, and Mayfield says so plainly: "A robot can cost $2 or $35 an hour, depending on who you ask." A range that wide means a buyer cannot judge a robot on a single price, and an investor cannot judge a robotics company on a single cost figure either. What matters is the gap between what the machine costs to run and what the work it replaces is worth, repeated reliably. Mayfield's newsletter stresses that shift from selling hardware to selling labour-as-a-service, with the emphasis on use and real-world reliability rather than the device itself. m:…9315

The money is still moving

The cheques have not stopped. Mayfield's roundup lists Google's $17 billion investment to build its first data centre in Wyoming, a $63 billion financing arrangement from a consortium of major financial institutions to support Anthropic's AI infrastructure through Broadcom, and Marvell's $115 billion agreement to acquire optical networking company Lumentum. NVIDIA and Oracle also announced plans to expand Vera Rubin AI infrastructure to multi-gigawatt scale. Taken together, the list shows capital still flowing into the largest bets in the sector, into data centres, chips and the infrastructure around them, even as the market turns more selective about what it will pay for. m:…9315

There are also signs of the work itself being done. OpenAI and ServiceNow expanded their partnership to bring AI agents into enterprise workflows, and OpenAI introduced GPT-6 Cyber, a specialised model for advanced cybersecurity research. In one collaboration cited by Mayfield, BigHat Biosciences used Claude to help design and test approximately 1,010 antibody variants in 10 weeks. That is the kind of result the new test asks for: not a promise about what a model might do, but a count of work completed and a time it took. m:…9315

The endowment view

Scott Wilson, chief investment officer of the Washington University in St. Louis endowment, is worth listening to because his record is one of concentration and patience. The endowment grew from just under $8 billion to nearly $20 billion under his tenure, and on taking the role he got rid of all 36 or 37 hedge funds in the portfolio. He says the private portfolio has compounded in the low to mid-20s over the last decade, against 10% for public equities, and that half of the portfolio sits in its top 50 look-through positions. That is a record built on backing a small number of positions hard, not on spreading bets thinly. m:…4116

Wilson says that if not for liquidity constraints, the endowment would probably target 70 to 80% in private markets, which shows how far his conviction in private investing runs. But he is equally clear about its limits. He calls the failure to add capital to winners the biggest mistake of his tenure, and he believes mega-funds will struggle to generate the historic multiples of invested capital they achieved on much smaller funds. Size, in his telling, is a drag on returns. The lesson for AI investors is direct: the discipline is not only picking well, but putting more behind the picks that prove themselves. m:…4116

The shakeout warning

Wilson's warning is the other half of the story. He argues that the current private-market valuation for AI companies is detached from fundamentals, and he predicts a shakeout in the AI industry as many companies lack defensible moats. A moat, in that telling, is whatever stops a rival doing the same thing at a lower price: proprietary data, a distribution lock, a cost edge, or work that is hard to reproduce. Without one, a high valuation is a claim on a future that competitors can undercut. That is the test Mayfield's labour-as-a-service advice points to as well: a robot that earns its keep has a reason to be paid; a model anyone can copy does not. m:…4116

The public market has already delivered its version of the warning. The All-In hosts' post-mortem puts the AI market sell-off at $4 trillion of market capitalisation wiped out. They also report that an OpenAI research paper disproved the exponential time hypothesis, and they warn that the United States is at risk of becoming a net importer of AI tokens if data centre development is blocked domestically. Data centre permitting, they say, will face significant uncertainty, likely stalling new projects by 2028. The money is still moving, the work is still being done, and the test has still changed: show the return, and show the moat. m:…2009