• 0 Posts
  • 3 Comments
Joined 1 year ago
Cake day: June 30th, 2024
  • rmean@feddit.orgtoTechnology@lemmy.world•OpenAI needs to raise at least $207bn by 2030 so it can continue to lose money, HSBC estimatesEnglish
    43·
    2 days ago

    OpenAI is a money pit with a website on top. That much we know already, but since OpenAI is a private company, there’s a lot of guesswork required when estimating the depth of the pit.

    HSBC’s US software and services team has today updated its OpenAI model to include the company’s $250bn rental of cloud compute from Microsoft, announced late in October, and its $38bn rental of cloud compute from Amazon announced less than a week later. The latest two deals add an extra four gigawatts of compute power to OpenAI’s requirements, bringing the contracted amount to 36 gigawatts.

    Based on a total cumulative deal value of up to $1.8tn, OpenAI is heading for a data centre rental bill of about $620bn a year — though only a third of the contracted power is expected to be online by the end of this decade.

    To check OpenAI ability to pay, HSBC’s team first had to build a model to forecast revenues.

    Its starting point is to put user numbers on an S-curve that by 2030 reaches 3bn, “equivalent to 44 per cent of the world’s adult population” ex China. That’s versus an estimated total user base last month of approximately 800mn:

    Advertising, agentic AI and possibly even Jony Ive’s thing can contribute to revenue by the end of the decade, For now, the business is mostly cajoling this user base to sign up for subscriptions.

    LLM subscriptions will become “as ubiquitous and useful as Microsoft 365”, HSBC says. It models that by 2030, 10 per cent of OpenAI users will be paying customers, versus an estimated 5 per cent currently.

    The team also assumes LLM companies will capture 2 per cent of the digital advertising market in revenue, from slightly more than zero currently.

    What results is gangbusters revenue growth:

    … but with a parallel rise in costs, meaning OpenAI is expected to still be subsidising its users well into next decade:

    … meaning each new OpenAI fundraise will be for shovelling cash to data centre owners:

    For what it’s worth, we can summarise a few of the assumptions HSBC is making for the estimates above:

    Total consumer AI revenue will be $129bn by 2030, of which $87bn comes from search and $24bn comes from advertising.

OpenAI’s consumer market share slips to 56 per cent by 2030, from around 71 per cent this year. Anthropic and xAI are both given market shares in the single digits, a mystery “others” is assigned 22 per cent, and Google is excluded entirely.

Enterprise AI will be generating $386bn in annual revenue by 2030, though OpenAI’s market share is set at 37 per cent from about 50 per cent currently. Everyone else stays more or less where they are now, market share wise.

    The bottom line is that, for OpenAI, it’s nowhere close to enough.

    HSBC’s model assumes that OpenAI’s rental costs will be a cumulative $792bn between the current year and 2030, rising to $1.4tn by 2033. The projection matches OpenAI’s eight-year guidance that CEO Sam Altman is exasperated at being asked to discuss.

    OpenAI’s cumulative free cash flow to 2030 may be about $282bn, it forecasts, while Nvidia’s promised cash injections and the disposal of AMD shares can bring in another $26bn. The broker also includes OpenAI’s $24bn of undrawn debt and equity facilities and, at the 2025 mid-year point, its $17.5bn of available liquidity.

    Squaring the first total off against the second leaves a $207bn funding hole, to which HSBC adds a $10bn cash buffer for safety’s sake.

    These estimates might prove overly cautious, though guessing how is a finger-in-the-air exercise.

    Each extra 500mn users OpenAI can grab will add about $36bn to cumulative revenue between now and 2030, while converting 20 per cent of the customers to paid subscriptions might bring in an additional $194bn over the same period, HSBC says. Assumptions for LLM spend and computing costs are flexed in similar ways, though the possibility of OpenAI chancing on Artificial General Intelligence is not put through the model.

    If revenue growth doesn’t exceed expectations and prospective investors turn cautious, OpenAI would need to make some hard decisions. Oracle has spooked debt markets, Microsoft’s support for OpenAI has been a bit flip-flop lately, and the next-biggest shareholder is SoftBank.

    The best worst option might be to call in some favours and walk away from data centre commitments, either before or at the usual contracted period of four to five years. HSBC says:

    Given the interlaced relationships between AI LLM, cloud, and chips companies, we see a case for some degree of flexibility at least from the larger players (less so for the neo clouds): less capacity would always be better than a liquidity crisis.

    What might not be clear from the above is that the HSBC software team is very, very bullish on AI as a concept. Here’s an entirely representative section of the note:

    We expect AI to penetrate every production process and every vertical, with a great potential for productivity gains at a global level. [ . . . ]

Some AI assets may be overvalued, some may be undervalued too. But eventually, a few incremental basis points of economic growth (productivity-driven) on a USD110trn+ world GDP could dwarf what is often seen as unreasonable capex spending at present.

    And when it’s put like that, is $207bn to tide things over really such a big ask?