Why the $25B generative AI mania might be a bursting bubble

When crypto took a nosedive in late 2022, venture capitalists were desperate for an escape hatch to park their idle cash. OpenAI launched ChatGPT right into that vacuum, scoring a million users in five days and sparking an unprecedented frenzy.

Over the first half of 2023, AI startups soaked up more than $25 billion in funding. But under the surface of massive headline checks, the sector is flashing classic signs of a speculative bubble.

The GPU black hole swallowing venture capital

According to PitchBook, 74% of surveyed venture investors backed at least one AI or machine learning startup, with 14% writing checks into more than six. Valuation tags went haywire: Inflection AI pulled in $1.3 billion at a $4 billion valuation, Anthropic raised $450 million, and Mistral AI locked down a $113 million seed round without a finished product or proof-of-concept.

So where is that $25 billion actually going? In normal software cycles, cash funds engineering headcount and user acquisition. In generative AI, funding heads almost straight into graphics processing units.

Startups are locked in bidding wars for hardware from Nvidia and AMD. Industry estimates peg global demand at roughly 432,000 H100 GPUs outside major Chinese tech firms. At roughly $35,000 a pop, that is roughly $15 billion earmarked purely for silicon chips. For context, GPT-4 required around 10,000 to 25,000 A100s, while Meta, Tesla, and Stability AI combined run on about 34,000 A100s. As Andrej Karpathy posted on X, tracking who gets how many H100s has become Silicon Valley's favorite parlor game.

The wrapper trap and fading novelty

While a handful of elite labs train frontier foundation models, thousands of newly funded startups are essentially glorified API wrappers. Companies like Jasper AI, which raised $125 million at a $1.5 billion valuation, rely on OpenAI pipelines to sell copywriting tools.

That creates an obvious vulnerability: zero proprietary intellectual property and zero leverage over suppliers. Now, the cracks are showing. Jasper recently laid off staff as enthusiasm cooled, and marketing startup Mutiny cut around 30% of its workforce.

Traffic metrics tell a similar story. ChatGPT experienced its first-ever drop in user interest, while rivals like Google Bard and Bing Chat have failed to capture meaningful runaway traction. Compounding the issue, a Stanford study found that GPT-3.5 and GPT-4 performance on multiple benchmark tasks has deteriorated significantly over time.

Why it matters

We think this critique nails the core structural flaw of the current wave. Capital is treating GPU procurement as if it were a sustainable business moat. In practice, startups are raising cash at exorbitant valuations just to pass it directly along to Nvidia and cloud hyperscalers, leaving themselves with thin margins and fragile software.

For builders and creators, the era of making easy money by wrapping an API endpoint in a shiny UI is wrapping up. Without proprietary data pipelines or deep enterprise workflow lock-in, wrapper startups remain completely exposed to platform shifts.

For the broader industry, hardware constraints won't prop up inflated valuations forever. Analysts expect the GPU supply crunch to balance out around late 2024. Once raw compute access becomes commoditized and free-flowing venture subsidies dry up, startups will actually have to generate sustainable profits. A few high-profile bankruptcies or brutal down-rounds might be all it takes to pop the rest of the balloon.