Why that AI model you love is "the fastest depreciating asset in human history"
Welcome to Cautious Optimism, a newsletter on tech, business, and power. Modestly upbeat.
Welcome to the week. Today in the newsletter, we have a brace of massive, new venture funds, stock market hilarity, notes on upcoming earnings and economic data, and a look at the value — or lack thereof — of cutting-edge LLMs. To work! — Alex
The rundown
📈 Trending Up: Latin American fintech … Iranian meddling in American elections … AI’s impact on election commentary? … crypto drama … tech money for Kamala … mobile gaming (BG) … Polymarket + Perplexity … phone bans in schools … AI coding help …
📉 Trending Down: Canadian real estate companies … Oyo’s private-market worth …
🤔 What Else?
Balderton Capital raises $1.3 billion: The FT reports that the fundraise is split between a $615 million early-stage fund and $685 million for later-stage bets. The new Balderton capital vehicles are an endorsement of European startup activity in today’s less-than-exuberant venture capital fundraising market.
TMTG is a joke: One of the most insane things on the public markets today is Trump Media and Technology Group’s market cap. The Truth Social parent company is worth about $5 billion this morning. That’s on par with Bill.com and Tenable. But in Q2 2024 its net sales fell from $1.2 million one year ago to just $837,000.
Sure, TMTG had a few million in interest income, but the company’s valuation appears to be a weird side-bet on the election? I don’t get it.
Stock market update: Shares were higher in Asia today, mostly up in Europe, and started the week lower in the United States.
📊 Upcoming data:
Earnings: Rumble and MariaDB on Monday, Nubank, Sea, Tencent Music, Ibotta, HUYA, and Serve Robotics on Tuesday, Cisco and TMC on Wednesday, Alibaba, JD.com, and Grab on Thursday.
Economic results: July producer price index (PPI), Core PPI on Tuesday, July consumer price index (CPI), Core CPI on Wednesday, initial jobless claims, two manufacturing surveys, retail sales, industrial production, and business inventories on Thursday, consumer sentiment and housing starts on Friday.
What’s that cutting-edge AI model worth?
Back in early July, CO covered a Goldman report that cast doubt on the near-term corporate utility of modern AI technologies. For critics of the current AI boom, the report was catnip. For AI belivers, it was useless doomerism.
After chewing on Goldman’s perspective, we dug into startups building AI models, focused on their ability to eat their own spend:
Put another way, all the money that AI model companies spend seems to accrue value only in their most recent models. This means that a massive amount of capital must be cycled through companies to ensure that they have built something of value instead of merely having invested in models that are now behind the leading edge of their market.
I am not sure if I precisely hit the mark, but my worry that well-funded startups are plowing equity (read: expensive) capital into AI model creation that the market, or the company itself will quickly deprecate, is shared by investors.
Pulling from a few talks that TWiST recently posted, here’s Atreides Management’s Gavin Baker on the topic:
A core belief of mine is if you're a foundation model company and you do not unique data — and internet scale distribution — you are the fastest depreciating asset in human history. And I think most of these companies are zeros and there's like 10 of them. And I think the only hope most of them have of getting [money] back is for Amazon to acquire them the way Microsoft did with Inflection [AI].
This is why Valor Equity Partners’ Antonio Gracias backed xAI, notably enough. From the same event, here’s Gracias discussing why Valor backed xAI and not other AI model companies:
We didn't invest in any model companies, any of the foundational model companies because we believe they were [basically] commodities. The models are commodities. The capacity to train them, the data centers are commodities. The only thing that's not really a commodity is the training data and the reinforcement learning.
That's it. So as we looked across the landscape […] if you weren't inside the big platform companies, you didn't have real proprietary data and really great reinforcement learning.
So why did he back xAI then? The underlying dataset. From later in the same riff:
If you believe what we believe, that these models are one of the most important technology innovations in human history, more akin to electricity than, say, the internet, then you want to make sure that they end up with the least amount of human bias.
And really, [the models] are imbued with the values of their creators. So if the people that are inserting the bias have a certain set of values, what those are, they will be inserted into the system. The X [Twitter] training data, we believe, has the least amount of bias in the system. Because it is a debate, a modern version of the Roman Forum, right? Where all sides are being heard, whether you like it or not. […]
That training data plus community notes makes X, we believe, the most effective training data for a system for an LLM[.]
Combining the Baker and Gracias arguments, AI models not built atop unique datasets will be worth nothing, and quickly. For companies like Anthropic et al that’s a worrying point to chew on. But there is a silver lining to consider.
If unique datasets are what allow AI models to stand out from their peers — allowing them to do more, and do it better — then the value of useful training data is massive. This means that data holders could be potentially worth more than the data modelers. As a person who makes data, that’s good news. For platform tech giants that have oceans of data, it’s even better news.
But for startups that don’t have a native dataset, or a few hundred billion to play with, it’s ill tidings.