OpenAI and Anthropic’s Valuation Risk: When Frontier Intelligence Stops Being Scarce

OpenAI and Anthropic just raised on a combined private valuation of about $1.8 trillion. Private investors are betting these labs stay the scarce source of frontier intelligence that customers will pay a fat premium for. AI demand may crush every aggressive forecast out there. Those valuations can still turn out completely wrong. If they do, the wipeout could look like the dot-com bust of 2000.

The bet is not that AI matters. Of course it matters. The bet is that the smartest models stay scarce enough and expensive enough for the companies that train them to collect enormous margins. That scarcity story is why OpenAI and Anthropic trade at insane multiples. Kimi K3, an open model from Chinese startup Moonshot, just made that story look much weaker.

Kimi K3 landed within striking distance of the best proprietary systems on nearly every serious benchmark. Artificial Analysis scores it within a few points of the top closed models. Moonshot said the full weights would come on July 27. Weights are the learned settings that make a model work. If the license is permissive enough, anyone can download it, optimize it, and run it on their own hardware. That severs who trained the intelligence from who gets paid each time it runs.

Think of a drug company spending billions to discover a molecule, then publishing the formula so any factory can manufacture it. The discovery still matters. Demand may explode. The money moves to manufacturing, distribution, and whoever ships at the lowest cost. That is the Kimi K3 signal. Model intelligence is becoming portable. Portable intelligence is very hard to price like a permanent monopoly. That is how OpenAI and Anthropic are priced today.

People collapse two trends into one and get confused. Training the absolute frontier is getting more expensive. Epoch AI estimates frontier training cost has been rising roughly 3.5 times per year. Reaching the tip still needs giant clusters, huge energy bills, and tens of billions of dollars. Inference at a fixed capability level is collapsing. Epoch’s data suggests those prices have been falling around 40 times per year. Frontier-level performance has historically reached consumer hardware in about eight months. Yesterday’s frontier becomes today’s cheap model. Then it becomes tomorrow’s open model on a workstation, in a car, or bolted into a robot. That second curve threatens the valuations.

OpenAI can spend at historic scale to ship the smartest model in July. That does not mean customers still pay a giant premium in January. Most buyers do not need the smartest model on Earth. A company reviewing invoices does not care that Anthropic’s Fable 5 just handled an obscure graduate math problem like the Jacobian conjecture. If another model does the invoices for one-tenth the cost, the math trophy is noise. A coding team may pay for the absolute best model on hard architecture work, then route tests, docs, and routine bugs to something much cheaper. Once several models clear a job’s capability threshold, more intelligence stops carrying the same premium. Competition shifts to reliability, speed, and cost.

There is a simple formula for this: useful intelligence divided by cost, multiplied by speed. A very smart model that is slow and expensive loses to a pretty smart model that is cheap and fast. OpenAI sees the same shift. It published a scorecard built around almost the exact phrase “useful intelligence per dollar.” Blunt admission: the company with the smartest model does not automatically win. The company that finishes useful work at the lowest total cost might.

Grok 4.5 from xAI makes that concrete. Artificial Analysis puts it on the cost-versus-performance Pareto frontier. Picture every model as a dot: intelligence up the side, cost across the bottom. You are on the frontier when nothing else is both smarter and cheaper at once. Grok 4.5 is not number one on every benchmark. It does not need to be. It charges $2 per million input tokens and $6 per million output tokens while staying near the intelligence frontier. Artificial Analysis measured about 31 cents per intelligence-index task. Close enough at a better economic point is where the market goes once models are good enough.

OpenAI’s last round valued it at $852 billion. The company says it generates roughly $2 billion in revenue per month. Anthropic’s latest round valued it at $965 billion on a pace above $47 billion a year. Extraordinary businesses. The valuations still assume extraordinary future value capture. The risk is not that people stop using AI. The risk is that model prices fall faster than usage can protect the profit. Every new model attacks the price of the models above it. Every open-weight release gives customers another exit from closed APIs. Every router sends easy work to the cheapest model that can finish it. Customers may spend more on AI overall while paying less for any one unit of intelligence.

OpenAI and Anthropic do have massive compute strategies. OpenAI has Stargate partnerships and a chip effort with Broadcom. Anthropic has committed more than $100 billion to AWS over ten years. Securing capacity through partners is not the same as owning the physical cost stack. Amazon owns the cloud. Nvidia owns critical silicon and software. Utilities sell the power. Data-center operators own the buildings. The lab pays all of them before it serves a customer. If the model stays scarce, the lab can still collect a premium larger than those input costs. If models converge and open weights spread, the premium gets competed away while the physical costs stay. That is the bubble: not a collapse in AI demand, a collapse in the assumption that a model lab remains the permanent premium provider of high intelligence.

Cheaper intelligence does not mean the world spends less on AI. The opposite is more likely. In the 1800s, better steam engines made coal more efficient. Britain burned more coal, not less, because cheaper mechanical work unlocked new factories and uses. That is Jevons paradox. Cut the cost of a useful agent by 90% and companies do not pocket the savings. They deploy thousands of agents to review every contract, test every build, and chase projects that used to be too expensive. Agents plan, act, inspect, and retry. A single task can burn millions of tokens before a human sees the answer. The token market can explode while value shifts away from any single model provider.

Who gets paid when tokens are produced anywhere? Nvidia is built for that world. It does not need to charge for every model. It needs the world to generate more tokens on Nvidia hardware. The company publishes NeMotron open models and open inference software like TensorRT-LLM to squeeze more work out of its GPUs. Giving away a capable model can make the chip underneath more valuable. That is commoditizing the complement: make the thing that drives demand for your main product cheap or free.

SpaceX is pushing an even more extreme version after acquiring xAI. It runs the Colossus data centers that train and serve Grok. It owns model ops, launch, terrestrial compute, and a giant satellite network. It plans deeper chip work with Tesla and Intel through TerraFab. The prospectus talks about orbital AI satellites as early as 2028. Engineering risks are huge. If it works, SpaceX owns an intelligence factory from energy to token delivery. Grok 4.5 already shows the early strategy: stay near the frontier and compete hard on useful work per dollar.

Physical AI weakens the model-lab story further. A self-driving car cannot send every camera frame to a distant data center. A humanoid robot cannot wait on the cloud before placing a foot. Factories and hospitals need local inference. Nvidia’s Jetson Thor runs large models on robots. Tesla cars have carried dedicated inference computers for years. As capable models get smaller and cheaper, intelligence spreads across millions of devices. The winner may be whoever embeds hardware, models, and software where intelligence turns into action.

OpenAI and Anthropic still have a real escape hatch. They are no longer just paid access points to models. OpenAI says ChatGPT has more than 900 million weekly users and over 50 million subscribers, with enterprise over 40% of revenue. Anthropic’s run-rate revenue has crossed $47 billion, and Claude Code plus enterprise products are becoming workflows, not interchangeable API calls. Distribution, trust, memory, integrations, and daily habit can sit above the model layer. Own the customer relationship and cheaper intelligence becomes an input, not a death sentence. Google never needed search queries to stay expensive. It needed to own where commercial intent showed up.

That escape route proves the larger point. These valuations cannot rest on model intelligence alone. They need distribution, workflows, infrastructure, or some other scarce layer that does not get copied every time a new model ships. The real AI bubble is not the belief that AI will transform the economy. That part is probably understated. The bubble is treating intelligence itself like a permanently scarce product. Kimi K3 is more evidence that scarcity is temporary. Grok 4.5 shows you can compete without sitting at number one. The cost of a fixed amount of intelligence keeps moving toward zero while consumption races toward infinity. Those trends can happen together. If they do, the largest fortunes will not necessarily belong to whoever trained the smartest chatbot this month. They will belong to whoever controls energy, chips, compute, distribution, workflows, real-world data, and the machines where intelligence becomes action. The models will keep changing. The factories that produce the tokens may be where durable value lives.

Check the video here.