All news

AlphaStar shows the AI frontier now runs on a parent's balance sheet

A small cobalt game piece alone on a graphite board resting on a massive pale stone block

DeepMind's agent beat two StarCraft professionals 10-1 while the lab lost £302 million in 2017. Frontier research now lives on corporate balance sheets.

Ten wins and one loss against two of the best professional StarCraft II players is the record DeepMind disclosed on 24 January for AlphaStar, its game-playing system. The ten wins came in December under professional match conditions. The loss came live, against a prototype trained for only a week and limited to a human player's camera view.

The match result is a research milestone; the bill behind it is the story for capital. StarCraft forces decisions over long horizons with most of the map hidden, the kind of problem logistics, trading and robotics face, and cracking it took compute and talent that few backers besides a corporate parent will fund without a customer. Frontier AI research has become a cost center inside a few technology groups. The companies investors can actually back will win by putting its published results to work, not by matching its spending.

Look at the training setup. DeepMind ran a league of competing agents for 14 days, with 16 of Google's v3 TPUs devoted to each agent, and each agent played up to 200 years of real-time StarCraft. The finished agent then runs on a single desktop graphics card. Almost the entire cost sits in discovery, before any product exists.

DeepMind's own accounts show who pays. The lab lost £302 million in 2017, as administrative expenses more than doubled to £334 million and staff costs nearly doubled to £201 million. Its turnover of £54 million came entirely from fees charged to other companies of its own group. A venture-backed lab with that profile would need a new round every few months, while a subsidiary of Alphabet needs only a parent willing to wait, as with Alphabet's other long bet, which has just started charging riders. The same parents set the price of research talent, which is why a seed-stage company rarely wins a bidding contest for a reinforcement learning scientist.

That leaves founders the more profitable half of the work. The methods behind AlphaStar, a transformer combined with a recurrent memory core and a league of agents that train against each other, are headed for a peer-reviewed journal, and the big labs publish faster than they turn research into products. Teams that carry those methods into scheduling, supply chains or game testing, with proprietary data from a real operation, capture the value a lab leaves on the table. Fanaura backs AI-first founders on that side of the ledger, the applied work our mandate covers.

A fair challenge is that corporate patience has limits. A lab funded this way can be cut or turned inward the moment its parent's earnings soften, which would slow the flow of methods to everyone else. DeepMind's 2018 accounts, due by the end of September, will show whether its group revenue is growing faster than its costs. A narrowing gap would mean the parent sees a return; a widening one would mean the research budget rests on a single board's goodwill.

The frontier will keep running on someone else's balance sheet, and the returns investors can reach sit with whoever ships what it publishes.

For founders

Taking published reinforcement learning methods into scheduling, supply chains or game testing? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.