Investors Ask What's Next As Foundation Model Mania Recedes
Plus, supply chain wreckage, and Google's antitrust troubles
At our November 2023 Cerebral Valley AI Summit, a top public markets investor came up to me and bragged that after our debut event in March he’d directed more than $1 billion to AI-related stocks like Nvidia. It had made his year.
I asked the investor if he was planning a similar trade based on our November event. He was skeptical. AI mania had already reached the broader world.
By the time of our New York conference this June, private market enthusiasm for foundation model investments had soured. People put a good face on it, saying they were excited about AI applications rather than models. But the signs were there that even in hardcore AI startup circles, some of the air was coming out of the balloon.
Taking an accurate temperature of AI investors’ mood matters a lot right now after a turbulent week on the stock market driven, in part, by worries that the AI play is losing its juice.
The public markets gave everyone in tech a scare. On Monday, the S&P 500 had its biggest single-day fall since September 2022. Tech stocks fared particularly poorly. A key volatility index spiked. On Thursday, though, stocks had a show of strength. As of now, the S&P 500 is up 3.5% over the past 5 days.
Still, some of the key technology stocks underpinning the AI boom are down meaningfully over the past month. Nvidia stock is down 21%. Alphabet is down 15%. Microsoft is off by 12%. Meta is down 3%.
Madeline and I spent the past week talking to private market investors, taking their temperature.
One top AI investor told me that for many months “every VC was saying the exact same thing.” AI investing had become a little too much like crypto. Everyone believed it was “going to the moon.”
Now investors are more uncertain. “The orthodox narrative is probably not exactly right. We don’t know what exactly is right. We’re in a liminal moment,” this investor said.
Rick Heitzmann at FirstMark Capital said, “People spent the last year or so saying AI, AI, AI. We’re going to try to invest in AI and pretend like that’s an unassailable strategy. It’s clear that’s not the case.”
Andreessen Horowitz started stockpiling GPUs, and doling out access to its startups.
There have been multiple warnings in recent months about irrational exuberance around AI. David Cahn at Sequoia wrote in June about AI’s $600 billion question, arguing that “the AI bubble is reaching a tipping point.” Goldman Sachs published a cautionary note. More recently, Elliott Management sent a letter to its investors asserting that megacap technology stocks like Nvidia are in “bubble land.”
It’s very possible that GPU stockpiling in Silicon Valley could soon come to an end. Cisco during the dot-com boom is one possible analogy: demand for networking gear plummeted once the crowd turned against internet stocks.
Here’s my read of where we are right now:
Venture capitalists no longer want to dump money into companies that are building their own foundation models. Adept and Inflection investors got out by the skin of their teeth with their novel licensing and talent deals. And sure, the Character AI deal was a win for investors, but it still looked much more like capitulation than victory. It’s now clear that Meta is planning to spend billions to keep its open source model Llama competitive. OpenAI, Anthropic, Microsoft, and a few others are relevant here — but the foundation model business is too expensive for even well-funded startups. And it remains to be seen how valuable foundation model companies are if they can’t develop popular products like ChatGPT. Models don’t seem to be valuable for long before they’re replaced by the next one.
How can you call this a bubble when it feels like we haven’t even found the core use case yet? ChatGPT is like the bitcoin of AI. It’s a popular use case that signals there’s something there. But all this infrastructure is being built on the belief that there’s something more. And unlike with crypto, I remain optimistic that large language models are going to spark valuable products in lots of domains. There’s been an Nvidia-specific bubble, because that’s what public market investors were able to bet on. But extraordinary run-up in its shares, which even after the recent pullback have more than doubled in a year, feels predicated on a GPU arms race continuing forever. Venture capitalists are eager to see startups build products with the AI capabilities they have today — not spend a fortune trying to build an indefensible tech edge for tomorrow.
One possibility is that some energy shifts to LLM-powered software. The quality of your foundation model isn’t the moat — product design, sales, user acquisition, and retention are the things that decide which companies win. Companies that provide data for AI seem buzzy. At the same time, artificial intelligence is helping startups generate new technology faster. Kane Hsieh at Root Ventures is an investor in Quilter, which uses AI to automate circuit board design and just raised $10 million in a round led by Benchmark. “We’re seeing a lot of new tech being invented,” Hsieh said. “In a boring way it’s back to venture: Figure out what’s going to change and how we can make money betting on it.”
The stock market turbulence is bad for the already dreadful tech IPO market. Bessemer’s Byron Deeter texted, “The already frigid IPO market may be even more cold,” adding, “The volatility is just feeding the IPO anxiety. The private markets have an incredible backlog of high quality companies right now, and yet very few are even considering going public in the next couple of quarters.”
I’m not calling a top on “AI” — but investing in this trend might not be so easy as buying a bunch of Nvidia stock and calling it a day. Machine learning has been a powerful force in technology for a long time now. That’s not going to change. Language models created a lightning in a bottle moment, persuading investors that we’d reached a real inflection point. I still think that’s true. But we’re definitely entering a moment where more applications need to emerge. As we wait for foundation models to improve by another leap, Silicon Valley is left with the hard work of building products that people want. I’m not losing the faith — we’ll be hosting another Cerebral Valley AI Summit in San Francisco on November 20. Revolutions don’t happen overnight.