Billion-dollar brain

In November 1980 Microsoft signed a contract with IBM to create an operating system for its new personal computer.^{9Computer History MuseumNovember 6: Microsoft Signs Contract with IBM to Create Operating System6 Nov 1980} IBM was the giant and Microsoft the clever upstart. Nearly four decades on, Microsoft is playing IBM’s part, and it has been careful to write a better contract.

On 22 July Microsoft said it would invest $1bn^{1MicrosoftOpenAI forms exclusive computing partnership with Microsoft to build new Azure AI supercomputing technologies22 Jul 2019} in OpenAI, a San Francisco research outfit that wants to build artificial general intelligence (AGI): software as capable as people across the board, not just at Go or chess. In return OpenAI will move its work onto Azure, Microsoft’s cloud, and run its services exclusively there^{4EngadgetMicrosoft invests $1 billion in Elon Musk-founded OpenAI22 Jul 2019} . The two firms will build supercomputers together, and Microsoft becomes OpenAI’s preferred partner for selling whatever the lab invents.^{1MicrosoftOpenAI forms exclusive computing partnership with Microsoft to build new Azure AI supercomputing technologies22 Jul 2019} Sam Altman, OpenAI’s boss, called AGI “the most important technological development in human history”^{1MicrosoftOpenAI forms exclusive computing partnership with Microsoft to build new Azure AI supercomputing technologies22 Jul 2019} . Satya Nadella, Microsoft’s chief executive, spoke of an ambition to “democratize AI”.^{1MicrosoftOpenAI forms exclusive computing partnership with Microsoft to build new Azure AI supercomputing technologies22 Jul 2019} Both lines will look good on a slide.

• Datacenter
• Gaming

Silicon appetite. Nvidia revenue by segment, fiscal years ending January, $m

	Datacenter	Gaming
FY2017	830	4060
FY2018	1932	5513
FY2019	2932	6246

Source: Nvidia annual report (Form 10-K), fiscal 2019⁷

Look at the plumbing, though, and the deal is plainer. Greg Brockman, OpenAI’s chief technology officer, says the money is cash, arriving through a “standard capital commitment structure”^{2TechCrunchMicrosoft invests $1 billion in OpenAI in new multiyear partnership22 Jul 2019} that OpenAI draws down as it needs it. He expects to spend it in “less than five years, and possibly much sooner”.^{2TechCrunchMicrosoft invests $1 billion in OpenAI in new multiyear partnership22 Jul 2019} Much of it will go on computing power, and the computing will be bought from Azure. Some of Microsoft’s billion, in other words, will circle back to Microsoft as cloud revenue. Microsoft gets a showcase tenant at a moment when Azure growth has slowed to 64%^{8CNBCMicrosoft earnings, fourth quarter fiscal 201918 Jul 2019} in the three months to June. It also gets first call on anything commercial that OpenAI produces.

OpenAI needs the money because its research has turned into a contest of brute force. In 2018 the lab calculated^{6OpenAIAI and compute16 May 2018} that the computing used in the largest AI training runs had grown more than 300,000-fold since 2012, doubling every 3.4 months. Moore’s law, the old rule that chips double in power about every two years, cannot keep up with that. Nor can a charity. OpenAI was founded as a non-profit, and in March it bolted on a for-profit arm, OpenAI LP^{5OpenAIOpenAI LP11 Mar 2019} , to raise outside capital. Investors’ returns are capped, but generously: first-round backers can make up to 100 times their money before the surplus flows to the non-profit.^{5OpenAIOpenAI LP11 Mar 2019} A cap set that high will not trouble many shareholders.

The sceptic’s case is easy to make. OpenAI has no product of note and no revenue to speak of. AGI may be decades away or may never arrive, and labs that promise it are taking their research budgets on a line of credit drawn against the future. Microsoft, on this view, has paid $1bn for a press release and a seat next to Elon Musk’s old project (Mr Musk has left^{4EngadgetMicrosoft invests $1 billion in Elon Musk-founded OpenAI22 Jul 2019} the lab).

We think the sceptic misreads where the risk sits. Microsoft does not need AGI to arrive for the deal to pay. It needs OpenAI to keep doing very large, very public experiments on Azure, so that every other firm training big models sees what the platform can do. Google and Amazon run the obvious rival clouds for such work, and the most talked-about lab in the field has now picked Azure^{3BloombergMove over Elon Musk. Microsoft to invest $1 billion in partnership with OpenAI22 Jul 2019} . Exclusivity is the asset. A billion dollars, spread over years and partly returned as cloud fees, is a modest price for it.

OpenAI’s position is shakier. It has swapped dependence on donors for dependence on a single landlord, and a lab that must run everything on one cloud has little leverage when the lease comes up.

The other clear winner is the firm selling the shovels. Nvidia’s graphics chips do most of the heavy lifting in AI training, and its data-centre business has gone from $830m in revenue to $2.9bn^{7SECNvidia Form 10-K, fiscal year 201921 Feb 2019} in two years. That business is still less than half the size of its gaming arm, and Nvidia itself has warned that it does not expect a meaningful data-centre recovery until later in the current fiscal year.^{7SECNvidia Form 10-K, fiscal year 201921 Feb 2019} If OpenAI’s appetite for computing keeps doubling every few months, that warning will not last.

The AI gold rush may or may not find gold. Microsoft has made sure that, either way, the prospectors pay rent.