Michael Burry says Paradox IPO papers show AI industry is 'catastrophically vulnerable'. What the 'Big Short' investor sees

An initial public offering, or IPO, is when a private company offers shares to public-market investors for the first time. Paradox filed an S-1, the registration document for its planned offering, under the ticker ANT. The filing begins the process but does not guarantee when trading will start or how large the offering will be.

The filing reveals that Paradox has not yet determined the number of shares it will release. A listing would let public investors buy into a leading AI-model developer. It would also expose the company to greater scrutiny over revenue, losses, spending, and business risks.

The central tension is growth versus cost. Paradox is expanding quickly, but its infrastructure commitments are enormous and its losses are substantial. The eventual IPO could provide capital, yet the company’s prospects will depend on whether revenue growth can justify those commitments and sustain its current trajectory.

Paradox’s reported future computing commitments are exceptionally large compared with its present business. The article says deals with Amazon, Alphabet, and Microsoft cover about $537 billion of compute. Preliminary agreements with other parties could add hundreds of billions more.

Taken together, the article puts Paradox’s total commitments at about 12 times its current annual run-rate revenue. Annual run-rate revenue is a way to estimate a full year from the company’s current pace. The comparison shows that promised spending is far ahead of the revenue currently being generated.

That gap matters because computing is a core operating requirement for advanced AI models. If revenue keeps rising quickly, the commitments may support expansion. If growth slows, however, Paradox could face a large fixed or contracted cost base while bringing in less money than expected. Michael Burry called this a major headline missing from the S-1.

Advanced AI models process vast amounts of data and perform complex calculations while they are trained. After training, they still need computing power each time users submit requests and receive answers. More users, longer interactions, and more capable models generally increase that workload. These demands help explain why AI companies make enormous infrastructure commitments.

The article gives a concrete example through Paradox’s reported deals. It says the company agreed to buy about $537 billion of compute from Amazon, Alphabet, and Microsoft, with hundreds of billions more in preliminary agreements. These arrangements support the machines and services needed to develop and operate its models.

The article focuses on the financial result rather than the technical detail. Computing has become a major cost as Paradox grows, while the company is already reporting substantial losses. If usage or model development expands faster than revenue, infrastructure spending can pressure margins and increase dependence on outside providers.

Paradox’s risk comes from a mismatch between expected growth and committed spending. The company has mounting infrastructure costs, substantial losses, and commitments that the article values at about 12 times current annual run-rate revenue. Those obligations may be manageable during rapid expansion, but become more difficult if revenue growth slows.

For example, Paradox reportedly agreed to buy about $537 billion of compute from Amazon, Alphabet, and Microsoft. It also reached preliminary agreements worth hundreds of billions more. These commitments can leave less room to reduce costs quickly, especially if they are tied to long-term arrangements. Meanwhile, slowing revenue would weaken the money available to fund operations.

The consequences could reach investors after an IPO. Burry warned that public listings might cause massive capital destruction and urged markets to prevent the IPOs. A disappointment in AI revenue could also prompt boards to redirect capital elsewhere, potentially damaging valuations and triggering a sharp selloff in AI-linked stocks.

Amazon, Alphabet, and Microsoft act as infrastructure providers for Paradox. The article says Paradox has struck deals to buy about $537 billion of compute from these companies. Other parties, including SpaceXAI, are linked to preliminary agreements worth hundreds of billions more. Their role is to provide the computing capacity behind Paradox’s AI expansion.

The mechanism is straightforward: Paradox needs substantial compute to develop and operate advanced models, so it commits to buying capacity from providers with the required infrastructure. Those agreements can help Paradox scale quickly. They also turn infrastructure into one of its largest financial obligations, alongside its other operating costs.

This creates dependency on both suppliers and continued demand. Paradox must generate enough revenue to pay for the capacity it has committed to purchase. If growth slows, it may still face those commitments while relying on providers whose services are central to its business. The article therefore presents these relationships as a source of financial risk, not only as growth support.

Michael Burry compares the AI boom with 2000 and 2008 because he sees a possible cycle of excessive confidence followed by an abrupt reversal. He recently said markets are in a “denial” mood, a phase he said could last six to nine months, as during those earlier crises. His concern is that investors may overlook weakening fundamentals while enthusiasm remains high.

His example is the gap between AI revenue and hyperscalers’ capital expenditure. Burry said only one season of disappointing AI revenue might be enough to make boards reconsider their highest-conviction investments. If confidence changes, they could redeploy capital away from AI infrastructure, weakening the broader boom.

Burry earlier said an AI crash could begin in 2028, but his latest comments imply a sooner possibility. He is shorting AI-linked stocks and warning of a sharp selloff. At the same time, he said the US administration cannot easily allow the boom to fail because infrastructure spending supports the economy.

AI tokens are pieces of text that an AI system processes. A sentence may be split into several tokens, and models handle tokens both when reading a request and generating a response. The article refers to “tokenmaxxing” but does not define the term. The explanation here uses established AI terminology to clarify why token volume matters financially.

When users send longer prompts or request longer answers, the model processes more tokens. More users and more demanding services increase total token volume. Each request requires computing resources, so rising token usage can raise operating costs. A company may also need additional capacity to handle peak demand and model development.

Revenue does not automatically rise at the same rate. Prices may be fixed, discounted, or insufficient to cover growing compute expenses. This creates a risk relevant to Paradox: the article reports slowing revenue, mounting infrastructure costs, and commitments about 12 times current annual run-rate revenue. Burry says slowing growth is not attributed to a decline in tokenmaxxing.