An IPO Makes the Financial Reality Harder to Hide

Going public changes the game because the company’s financial position becomes much more visible. That matters when you’re spending huge amounts of money on compute and training while still trying to establish a path to profitability. Delaying an IPO buys more time before public investors can examine those economics in detail. The official explanation around safety may be genuine, but safety alone doesn’t explain the enormous financial pressure surrounding these companies.

Training the next generation of models requires massive amounts of computing power, and pushing performance higher can require dramatically more spending for relatively small improvements. If you’re spending ten or even a hundred times more to squeeze out another few percentage points of performance, eventually you have to ask whether customers will actually pay enough to make that investment worthwhile. That’s the financial question sitting underneath the AI boom.

LLM Prices Are Falling Fast

The biggest threat to the current AI business model may not be another frontier laboratory. It may be commoditization. LLM prices are falling extremely quickly, much faster than many previous technology markets experienced. Open-weight models are contributing to that pressure because customers increasingly have alternatives that are capable enough for real-world workloads.

The result is a market where capability continues to increase while the cost of using that capability continues to fall. That’s great for customers, but it creates a much harder environment for companies whose valuations depend on maintaining enormous revenue growth from proprietary models. When the product gets cheaper every year, the company needs either massive scale or another source of economic value. Simply having the most expensive model isn’t enough.

Open-Weight Models Are Changing the Game

Open-weight AI makes the problem even more obvious. If a company can take an open model, run it in its own environment and get comparable results for a fraction of the cost, there is very little reason to keep paying premium prices for a proprietary model on every workload. OpenRouter data discussed here shows proprietary model share falling from roughly 60% to 25%.

AT&T provides another example, moving AI workloads from proprietary models toward cheaper open alternatives and reporting savings of up to 80%. That’s the kind of shift that matters. Enterprise customers don’t necessarily care about the logo behind a model. They care about whether it performs well, integrates with their systems and costs less. If an open model can do the job, businesses have a strong incentive to use it.

Meta Has a Completely Different Advantage

This is where Meta becomes one of the most interesting companies in the entire AI race. Meta is already a massive public company with an established revenue engine. It isn’t dependent on AI revenue to justify its existence. That gives Meta an advantage that OpenAI and Anthropic don’t have. Meta can spend aggressively on AI using money generated by its existing businesses.

It can build models, improve them, distribute them across its platforms and even give them away without needing every AI interaction to generate direct revenue. Meta’s Muse Spark 1.3 is presented as competitive with leading AI models on benchmarks while being available for free. If the model is good enough for developers and businesses to actually use, Meta doesn’t necessarily need to charge for every token. It can use AI to strengthen an ecosystem that already generates enormous amounts of revenue.

Meta Can Afford to Play the Long Game

Meta has already demonstrated that it is willing to spend heavily on ambitious technology. The company’s previous metaverse strategy may have looked questionable, but the same willingness to invest aggressively becomes a very different advantage when applied to AI.

The competition is also much broader than OpenAI versus Anthropic. Google has enormous resources and infrastructure. Meta has enormous resources and distribution. SpaceX and xAI are pursuing their own long-term strategies. The companies competing in AI don’t all have the same financial constraints. Meta can afford to experiment, release models, improve them, distribute them to millions of users and continue spending without requiring the AI business itself to immediately generate enough revenue to support a trillion-dollar valuation. That makes Meta a potential dark horse.

The Trillion-Dollar Comparison Gets Stranger

Consider the comparison with Walmart. Anthropic is being discussed around a $965 billion valuation, while Walmart is around $870 billion in the numbers being used here. Anthropic’s revenue is presented at roughly $65 billion, while Walmart generates around $713 billion. The comparison is striking because Walmart is an enormous revenue-generating business with a physical infrastructure and a global retail operation, while Anthropic is being valued at a similar or significantly higher level despite having a fraction of the revenue.

Anthropic’s annualized revenue pace is also described as increasing from roughly $47 billion to $65 billion. That’s impressive growth, but rapid growth doesn’t automatically make a $2 trillion valuation reasonable. Eventually, the underlying economics have to support the number. That’s the part that deserves scrutiny.

Compute Costs Are Exploding

The spending problem doesn’t stop with model training. Anthropic needs massive amounts of compute to keep competing, and the cost of that infrastructure is becoming one of the biggest issues in the entire AI industry. A potential Anthropic compute arrangement discussed here could cost as much as $500 billion, compared with roughly $180 billion previously discussed through 2029.

Those are extraordinary numbers. The basic problem is straightforward. Better models require more compute. More compute requires more money. At the same time, competition and open-weight alternatives are pushing the price of AI downward. If your costs keep increasing while the price customers are willing to pay keeps decreasing, you eventually have to prove that your scale and margins can overcome that gap.

Safety or Financial Pressure?

AI safety has become a major part of the public conversation, but the timing of these IPO decisions raises a different question. Safety isn’t a new concern. People have been warning about the risks of advanced AI for years. The issue is whether safety actually explains the decision to delay going public.

There is a much simpler financial explanation sitting right in front of everyone. These companies are spending enormous amounts of money, they need increasingly large amounts of compute, and the economics of selling AI are becoming more difficult as models get cheaper and open alternatives improve. That doesn’t prove that safety concerns are fake. It does mean the financial explanation deserves serious attention instead of being dismissed.

The AI Race Isn’t Just About Model Quality

The AI race is often framed as a competition over who can build the smartest model. But that misses an important part of the story. The company that wins may not be the one with the most impressive benchmark score. It could be the company that can continue spending when the market becomes more competitive and prices continue falling.

That is one reason Meta’s position is so interesting. It already has massive revenue, enormous distribution and the ability to fund AI development from businesses that existed long before generative AI. OpenAI and Anthropic have to build enormous AI businesses while simultaneously spending enormous amounts to keep their models ahead. Meta doesn’t necessarily have to play by those rules.

Businesses Shouldn’t Depend Entirely on Someone Else’s AI

There is also a lesson here for companies building their own software. AI shouldn’t simply be bolted onto an existing product as another API dependency. Businesses need to think about who owns the infrastructure, where the data lives, how the models are accessed and what happens when a vendor changes its pricing or availability.

That is the thinking behind StartupHakk’s approach to custom software, database architecture, API design, system integration and scalable infrastructure. AI works best when it is treated as part of the architecture rather than a thin wrapper around someone else’s API. OpenMonoAgent.ai is another example of that philosophy. It is an open-source, terminal-native AI coding agent designed to run entirely on local LLMs, with zero API costs and zero telemetry. The broader idea is simple: own what you can, control what matters and avoid building your entire business around somebody else’s pricing and infrastructure.

Conclusion

The AI market is entering a much harder phase. The technology is getting better, but better technology does not automatically mean better economics. Models are becoming cheaper, open-weight alternatives are improving, enterprise customers are looking for savings, and the amount of money required to train the next generation of frontier models keeps increasing.

That’s why the valuations of OpenAI and Anthropic deserve serious scrutiny. The technology can be incredible and the valuations can still be ridiculous. Revenue can be growing rapidly while the cost of staying competitive grows even faster. Meanwhile, Meta has a completely different advantage. It already has a huge revenue-generating business, massive distribution and the ability to spend aggressively on AI without requiring AI revenue alone to justify the entire company.

That’s why Meta could be the dark horse. The companies that ultimately win the AI race won’t necessarily be the ones charging the most for AI today. They may be the ones that can afford to keep playing when AI becomes cheaper, more open and more commoditized. For businesses, the same lesson applies: this is why having the right Fractional CTO strategy matters. Instead of blindly depending on AI vendors and their changing pricing or infrastructure, companies need technology leadership that helps them own, control and properly integrate AI into the architecture of the business. And that is exactly why the current OpenAI and Anthropic valuations deserve a much harder look.