The Trend Points One Way

Local AI wins in the long run. Not this month and not for every task, but the trend only goes one way. The price of tokens keeps dropping and the open models keep getting closer to the top ones. Cheap models are already 80 to 95 percent as good. With a good harness and solid software engineering around them, they are every bit as good for email, tickets, most coding, and a lot of other everyday work.

The expensive labs need the price to stay high because they are spending billions training the next model. The open model does not carry that bill. It already works today. Buyers are done waiting. Companies want their own compute, their own models, and their own data. They are not just talking about it. They are doing it.

Revolut Cut Its Bill to a Tenth

Revolut is Europe’s most valuable startup, valued at 200 billion. At the start of this year it leaned on the frontier labs. Its 2,000 engineers used cloud coding tools, and some burned $1,000 to $10,000 a month. The bills kept climbing and the returns were not there.

Revolut switched its default to an open-source model called GLM. All it needs is GPUs, which it rents from Together AI. By Revolut’s own count, the bill is about a tenth of what it paid OpenAI and Anthropic. Engineers get unlimited tokens instead of spending caps. Coding time is down 60 percent and lead time is down 70 percent. Revolut still uses cloud and GPT, but only as the exception for tasks those models do best. Everything else runs on open-source.

How many use cases can companies still really justify sending to frontier models? Very few. Plenty of companies will route everything to OpenAI and Anthropic today, but for how long?

The Steepest Commodity Curve

Chinese models run 112 times cheaper than Anthropic for every million tokens. Anthropic runs at $26 or $56 depending on the list, Meta at $1.50, X and Google at $1, and Chinese models at 50 cents. The gap is the steepest commodity curve in the recorded history of technology. Oil needed 40 years to compress that way. Semiconductors needed 20 years. AI is getting there in months.

Those higher prices look careless. They are a gamble that trust, safety, and enterprise ties will keep the markup in place long enough.

AT&T Routes Most Tokens Through Open Models

AT&T is betting on open AI models to control generative AI costs. The operating shift is moving older AI use cases from closed models to open alternatives. Smart model routing is central to the strategy, with cash-aware routing deciding when to use open models or premier options. AT&T serves 65 to 75 percent of its tokens through open models and cut the cost of coding in some areas by as much as 56 percent. This is exactly what the big frontier companies do not want you to do.

Models Have Been Oversold

Palantir CEO Alex Karp says every enterprise he deals with is livid. Customers are paying for tokens that create no value. The models have been completely and irresponsibly oversold. He cannot safely give them to an enterprise without being certain that the alpha of the business will not transfer to the model tomorrow. His message to investors is clear: pick up the phone and call a CEO in private. Call two or three. They are twice as livid.

Running Frontier-Level Models on Ordinary Hardware

Huge models are already being served on hardware that costs $2,000 to $4,000. This is not $10,000 equipment. Any business can buy it and run very large-scale models. A six-year-old 3090 with 24 GB of VRAM still runs incredibly well. Start with one for development; add a second for redundancy if you are a business. Pair it with a machine that has 64 GB of RAM (DDR4 is fine).

That setup comfortably serves 60 tokens per second on a 180-billion-parameter model such as Quen 3.8 flash next. The model sits within five to seven benchmark points of AI4. You are running some of the top frontier models on something that costs a few thousand dollars to build.

Own the Infrastructure Instead of Renting Tokens

AI should not have a meter. It should not be a subscription you rent. It should be infrastructure you own, sitting on your desk, serving your code, and answering only to you. Installation can be as simple as running a curl script, selecting the option that auto-detects hardware, then setting up a secure tunnel with an email OTP. The architecture keeps heavy local GPU inference completely separate from the coding agent via a secure relay. On the client side you install the VS Code extension, connect the relay, and the local setup is accessible from anywhere.

Data sovereignty is another reason the shift is happening. You keep control over your own data. It never leaves your infrastructure. When you call out to OpenAI and Anthropic you are pumping customer data out to the cloud. There is a reasonable prediction that the providers have been training on that data all along, regardless of any boxes checked.

Most Companies Have a Leadership Problem, Not a Technology Problem

Most companies do not have a technology problem. They have a leadership problem, and they are paying for it in missed deadlines, failed integrations, and AI investments that deliver nothing. The businesses winning right now are not the ones chasing the latest AI trends. They are the ones that built on solid engineering and treated AI as infrastructure they own, control, and integrate into software that actually works.

Custom software teams that design AI into the architecture—running in the customer’s own environment, not a vendor’s cloud and not on someone else’s pricing schedule—are already delivering that model. An open-source platform that provides a terminal-native AI coding agent running entirely on local LLMs demonstrates the approach: zero API costs, zero telemetry, full ownership. It is growing because serious engineers recognize real infrastructure when they see it.

Fractional CTO leadership can deliver the same standard without the full-time executive cost: strategic architecture, hands-on delivery, and clear accountability instead of slide decks that gather dust. When organizations are ready for custom software built to last, with AI integrated where it actually makes sense, the path is to treat it as true infrastructure they own.

The prices are collapsing. The open models are already close enough for most work. Enterprises with real invoices are switching. The math is straightforward. The only remaining question is how long any given company keeps paying the higher rate.

Conclusion

Token prices continue to fall sharply, with Chinese open models running as much as 112 times cheaper than Anthropic’s $56 per million tokens. At the same time, open models have already closed most of the quality gap for everyday work such as email, tickets, and the majority of coding tasks. Companies facing real invoices are responding: Revolut cut its AI bill by roughly 90 percent, gave engineers unlimited tokens, and reduced coding time by 60 percent and lead time by 70 percent, while AT&T now routes 65 to 75 percent of its tokens through open models.

Enterprises are also growing frustrated with paying for tokens that deliver little value, and many are choosing their own compute, their own models, and their own data. Affordable hardware in the $2,000 to $4,000 range can already serve large models at useful speeds, giving organizations full control and keeping data inside their own infrastructure. Local AI does not win every task this month, yet the direction is unmistakable. Prices keep dropping, open models keep improving, and the companies that treat AI as infrastructure they own rather than a metered service are already capturing the advantage.