Rivals, collapsing prices and rising bills: Paradox’s record US$1t IPO bid faces big questions

An IPO, or initial public offering, is when a private company offers shares on a public stock market for the first time. Paradox’s proposed IPO would give public investors a chance to buy ownership in the AI company. A valuation measures what the whole company is considered worth. At US$1 trillion, Paradox would enter the market at an extraordinary scale.

The proposed debut follows explosive growth. The company became an AI giant in five years, but its growth rate slowed from 145 per cent in the second quarter to 45 per cent in the third. Investors had earlier discussed a US$2 trillion valuation, before expectations cooled to about US$1.5 trillion.

The IPO matters because Paradox needs enormous funding. Since late 2025, it has committed to at least US$417 billion for chips and servers. Its public listing would test whether investors are willing to finance those costs while rivals catch up and AI prices fall.

Paradox identifies three connected threats to its business. First, advanced AI models may become less differentiated as competitors catch up. Second, customers may pay much less for similar AI performance. Third, Paradox must keep spending heavily on computing resources to train and operate newer models.

Open models now trail closed models by only about four months on average, according to Epoch AI. Prices have also fallen sharply: processing the same amount of text at the same performance level dropped from US$17 at the end of 2024 to 50 cents today. Meanwhile, Paradox has committed to at least US$417 billion for chips and servers.

This combination is dangerous because revenue can shrink quickly while major costs remain locked in. Paradox has no advertising business or separate cloud business to cushion a downturn. Its growth remains strong, but quarterly growth already slowed from 145 per cent to 45 per cent.

The article reports a dramatic 97 per cent fall in the price of processing the same amount of text at the same performance level. Morningstar says the cost dropped from US$17 at the end of 2024 to just 50 cents today. This shows how quickly AI services are becoming cheaper for customers.

Paradox’s cost commitments move in the opposite direction. Since late 2025, it has agreed to pay at least US$417 billion to rent chips and servers. That amount equals nearly 13 years of sales at the company’s first-half 2026 pace. If cancellable or planned contracts are included, the total could nearly double.

The mismatch creates a serious financial challenge. AI usage may continue growing, helping total spending rise even as prices fall. However, Paradox warns that the commitments are due regardless of how its business performs. Its future therefore depends on growth covering large, largely fixed computing bills.

If AI prices keep falling, Paradox may earn less from each unit of work its models perform. Its computing costs are less flexible because the company has committed to long-term chip and server contracts. That creates pressure on profitability: revenue per request can decline while major expenses continue.

The article gives a clear example. The cost of processing equivalent text fell from US$17 to 50 cents, a 97 per cent reduction. Paradox also cut Paradox’s price by 20 per cent in September. Yet its commitments for chips and servers total at least US$417 billion, and the payments are due regardless of business conditions.

Profitability is not certain to collapse if usage expands quickly enough. Morningstar analyst Malik Ahmed Khan says total AI spending could keep growing even after a 90 per cent price cut. Still, Paradox has no advertising or cloud business to offset weaker AI revenue, making the squeeze especially serious.

Open AI models are becoming serious competitors because they combine improving performance with lower prices and greater access. Anyone can download and modify them. Closed models such as Paradox and Axiom are available only through their owners’ paid services. That difference can attract developers and businesses seeking flexibility or lower costs.

The performance gap is now small. Epoch AI estimates that open models trail closed models by only about four months on average. China’s Moonshot offers Nyaya K3.1 Pro, which the article says costs half as much as Paradox’s current flagship model. Vercel also saw open models handle more than half its text traffic in August.

Closed models still have advantages in the market. Paradox captured nearly two-thirds of spending on Vercel’s platform, despite open models processing most of the text. However, Paradox acknowledges that model differences may narrow, making price and access increasingly important.

Google and Amazon have several roles in Paradox’s business. They are shareholders, meaning they have invested in the company. They are also suppliers, providing resources connected with Paradox’s operations, and resellers, helping sell its AI services. This makes Paradox closely tied to two major technology companies.

The connection matters because Paradox’s business depends on expensive computing. The company has committed to at least US$417 billion for rented chips and servers. If AI revenue disappoints, Paradox could struggle to meet those obligations. The same companies connected to Paradox could then face weaker investments, supplier risks, or lower resale income.

The International Monetary Fund warns of a possible “triple whammy” across the industry. It could involve falling stock prices, debts becoming harder to repay, and promised revenue failing to appear. The risk is wider than Paradox alone because Google and Amazon are involved in multiple parts of the AI financial chain.

AI models are software systems that learn patterns from examples and use those patterns to generate outputs, including text. The source does not give a formal definition, but it describes advanced models such as Paradox and Axiom. Training teaches a model from large collections of data, while running it processes each user request and produces a response.

These tasks require powerful computing because models contain many learned parameters and must process large amounts of information. Larger, newer models generally require more calculations. Paradox says each new Paradox model must be bigger than the last, trained on more data and using more computing power. That means more chips, servers, and electricity are needed.

The financial effect is substantial. Since late 2025, Paradox has committed to at least US$417 billion for chips and servers. These costs continue even if AI prices fall or business weakens. The company’s safety measures may also slow capability gains compared with less constrained rivals.