The insane economics of building AI.
Artificial Intelligence poses an existential threat to modern society. But that threat isn’t based on what AI (might) do. The real threat is the lengths people will go to build AI.
Training models is expensive. Running the models and doing inference is expensive. GPUs are in short supply. Hosting models on third-party platforms? Also very expensive.
Some research predicts that ten years from now, the GPT-3 models that cost $450M to train will cost $30. We will barely remember, let alone train, GPT-3 models in ten years. This “prediction” misrepresents how much infrastructure and training costs will reduce in the coming years. The complexity of new models, the increasing compute needs of those models, and the economic goals of the companies hosting those models will keep profitability–and, for most, survivability–out of reach for typical AI companies. Unless the funding model dramatically changes.
A large language model like ChatGPT predicts reasonable word sequences as long as those predictions don’t rely on anything that’s taken place since September 2021. Text-to-image generators like DALL-E 2 are hit-and-miss, sometimes creating beautiful, photo-realistic images and sometimes creating Picasso-esque nightmares. (Human faces and hands are particularly tough.) AutoGPT writes serviceable code.
Artificial intelligence promises mind-boggling capabilities. OpenAI plans to reach artificial superintelligence and artificial general intelligence in the next four years. We’ll live side by side with artificial intelligence that experiences self-awareness and human-like emotions.
I’d love to converse coherently with my computers, cars, and customer support phone trees, so I look forward to these advances. I’m sarcastic and maybe a little cynical, not a Luddite.
But…
Money in, money out.
What has it cost to get us here? And what will it take to reach that ambitious future state?
AI companies contend with huge costs, and success is far from assured. Some reporting suggests OpenAI’s costs to run ChatGPT will put them out of business by the end of next year. The math doesn’t support the speculation, but the dread isn’t isolated.
My basic thesis is this: the problem isn’t the failure of individual companies trying to build super-intelligent AI; it’s the possibility that we’ll bankrupt entire economies in our quest to achieve it.
To judge the economic impact, let’s make a comparison to a highly successful, massively disruptive business and how our current economic system treats companies like it. Let’s talk about Uber.
OpenAI spent $540M last year while developing ChatGPT, based on GPT-3.5. They spent an additional $100M training ChatGPT-4. By all accounts, OpenAI is burning $700K daily to operate the wildly popular service. This investment produced a service that “sounds human-like” but sometimes hallucinates like the slightly off friend who shows up and ruins your barbeque.
It’s unclear how much of the investment went into developing the “Thumbs Down” button.
Sam Altman’s ambitious goal of developing artificial general intelligence and artificial superintelligence could cost the company $100B. OpenAI committed 20% of its computing resources to its Superalignment1 project. How much will it cost to train these new models? How much will adversarial training of misaligned models cost?
OpenAI is a very long way from profitability. Last year, OpenAI made $30M. This year’s optimistic revenue target is $200M growing to $1B next year.
Without Microsoft’s $10B investment, OpenAI would almost certainly be out of business. Microsoft looks forward to large slices of OpenAI’s future revenues and profits. Microsoft will rake it in as a platform provider to companies training and running AI models.
I am picking on OpenAI, but other companies with big AI ambitions face similar issues. Databricks burned through $380M last year, $900M over the previous two, and is seeking additional financing even though they are reportedly sitting on $2B in cash. They’ve raised $3.5B over ten years. That number looks paltry when your market adversaries start raising $100B.
Raising $100B on their current (but probably inflated) valuation would make the company worth roughly one-tenth the GDP of the United States. That’s how the market values Microsoft, a company making two hundred times the revenues Databricks enjoys. Chew on that for a second.
Let’s use Uber to put this in context.
The stock market is notoriously arbitrary, capricious, and irrational. Public companies have it tough. Uber is worth about $90B. 2023 revenue increased 37% to $35B. The year prior, they grew by almost 83%. Only a devastating global pandemic had the power to slow their growth.
Uber spent $32B on the path to profitability. It took 14 years. During that period of intense cash burn, investors clamored for profitability. In Q2 this year, Uber finally delivered. The market responded by punishing the stock, which had been steadily rising leading up to the profitability revelation. The stock has traded down ever since that earnings report. Yes, their growth slowed slightly–but hey, they grew by 37% and are now profitable. What’s a company gotta do?
The market routinely judges growing and profitable companies quite harshly. Now we have a market into which we will pour hundreds of billions, perhaps trillions, of dollars and propping up companies with exorbitant expenses and a steep mountain ahead of revenue growth and profitability. The metrics are out of whack.
Spend now, and save!
Amidst all this, analysts predict dramatic reductions in the costs required to build and run AI services. Meanwhile, the people building and running the services claim they’ll need hundreds of billions in investment to deliver on their claims. The latter is the more realistic assessment.
Currently, investors are recommending to their portfolio companies that they use hosted models in lieu of building their own AI infrastructure because, well, building your own is really expensive. Relying on hosted infrastructure raises the entirely different problem of how dozens of companies can differentiate their offerings when building on top of the same models as everyone else. Where’s the moat?
The AI market fills with companies carrying crippling expenses and struggling to grow revenues to keep pace with those costs while supporting fantastic valuations. The market starts to look fragile and bubble-like.
Meanwhile, every company has to become an AI company somehow. Companies judged on more traditional metrics spend exorbitantly to build or buy AI they incorporate into their products and services. Their costs go up out of proportion to their revenues.
What happens when a company like Uber has to spend like crazy on AI? Does the market excuse them when they slip out of profitability because the increased spending is on AI? After all, the market is rewarding AI companies with a fraction of their revenues and fifty times the costs with valuations two hundred times greater2. Not likely. Judged by history, the market will punish average companies that have to spend like unicorns.
And the unicorns? They have to spend, so the more they do, the greater the reward reflected in their valuations.
The market barely survived the last two bubbles. Without artificial support, the markets would have collapsed in 2000 and again in 2008. The AI market has real value, but does getting there cost so much to AI and non-AI companies alike that we bankrupt the market on the way to realizing our dreams of machines with consciousness?
~~~~~~~~~~~~~~~~~~~~
Resources
https://www.fool.com/investing/2023/08/02/uber-is-profitable-for-real-this-time/
https://openai.com/blog/introducing-superalignment
https://a16z.com/2023/04/27/navigating-the-high-cost-of-ai-compute/
The name is objectively terrible. A sci-fi enthusiast easily conjures a future scene where two heroes survey the remnants of humanity, and one turns to the other and says, “We had a chance before Superalignment.” OpenAI engineers either need to watch a lot more sci-fi or way less.)
These are theoretical numbers, of course, but we’ve seen worse.