In January 2025, one cheap Chinese model wiped roughly $1 trillion off US markets in a single day. That was the warning shot. What’s followed is a grinding price war that’s pulling real customers toward cheaper models right now, while the expensive US labs behind the frontier carry heavy debt against thin revenue. For construction, it cuts both ways: the AI inside your tools gets cheaper to run, but the ground under your vendors and your funding gets less stable.
On a single Monday in January 2025, a little-known lab called DeepSeek released a model that worked nearly as well as America’s best, at a fraction of the cost. By the close of trading, about $1 trillion had been erased from US markets. Nvidia alone lost roughly $589 billion that day, the biggest one-day drop for any company in US history.
Nvidia clawed most of it back within weeks. But the lesson stuck: one cheap model, out of nowhere, could crack the confidence holding up the entire AI trade. What’s happened since is slower and, for the US AI industry, more dangerous. The shock has hardened into a price war, and cheaper models, mostly open-weight and mostly Chinese, are now good enough for everyday work that big companies are rerouting real budgets to them.
This lands on construction for a plain reason. The models under pressure are the same ones now baked into your software, and the capital funding the AI boom overlaps with the capital funding ConTech. When the economics shift, both feel it.
Cheaper AI is good news for anyone buying construction tools and bad news for anyone raising or backing ConTech. Both things are true at once, and both trace back to the same price war.
The cost of AI is collapsing
This isn’t a forecast. Companies are already switching.
The clearest signal came from Coinbase. Its CEO said the exchange cut internal AI spend by nearly half by defaulting its engineers to open-weight models like Zhipu’s GLM and Moonshot’s Kimi, while token usage actually went up. No usage caps, no rationing, just a cheaper default sitting in front of routine work. The startup Lindy went further and moved off Anthropic’s Claude entirely to a Chinese model to stop its AI bill from outrunning payroll.
The reason is math. Open-weight models run roughly five times cheaper than the top US models on many tasks, while landing within a few points of them on standard coding benchmarks. When quality is close and price is a fraction, procurement makes the call for you. And the pressure was building anyway: Uber burned through its full-year AI coding budget in about four months, and Meta warned staff of an “exponential” rise in AI spend.
- Frontier pricing hurt. Coding assistants generate huge token volumes, so even small per-token gaps turn into big bills at scale.
- The cheap options got good. Open-weight models arrived close enough on everyday work that the quality argument stopped holding.
- You can self-host them. Open weights mean a company can run the model on its own servers, keep its data in-house, and skip the vendor markup entirely.
Why the cheap models showed up so fast
US labs say their work was copied, and it’s now a fight in Washington
Part of the reason the cheaper models caught up so quickly is a technique called distillation: training a smaller, cheaper model on the outputs of a bigger, expensive one. Done to your own model, it’s standard practice. The dispute is about doing it to a rival’s model you were never licensed to use.
In February 2026, Anthropic accused three Chinese labs, DeepSeek, Moonshot, and MiniMax, of “industrial-scale” distillation, generating more than 16 million exchanges with Claude through roughly 24,000 fraudulent accounts. In June it escalated, telling the Senate Banking Committee that operators tied to Alibaba’s Qwen lab ran the largest effort yet, about 28.8 million exchanges through nearly 25,000 fake accounts. Alibaba denies it. The accusations have since fed into export-control fights and proposed sanctions.
The takeaway for anyone buying AI isn’t to pick a side in a geopolitical spat. It’s that the price gap you’re seeing is partly a real efficiency story and partly the product of a fight over who paid to build these capabilities in the first place. That fight is not close to over, and it can reshape which models are legally available to you.
The expensive side is heavy with debt
The frontier buildout only works if revenue shows up fast
While buyers hunt for cheaper models, the US labs and cloud providers are spending at a scale that assumes enormous future revenue. Oracle just posted its worst week on Wall Street since the 2001 dot-com bust, dropping about 19% in five days as investors fixated on the debt behind its AI buildout. The DeepSeek scare showed how fast confidence can snap. The debt is what makes the next snap harder to survive.
There’s a demand problem underneath it too. An MIT study found 95% of enterprise generative AI pilots produced no measurable impact on profit and loss, though that figure covers pilots and its method has been argued over. And investor Michael Burry has hammered the “circular” deals propping up the sector, where Nvidia invests billions in OpenAI, then sells OpenAI the chips it buys with that money, and also owns a slice of the cloud firm renting those chips out. Nvidia calls the concern overblown. Treat it as a live argument, not a settled verdict, but it’s the argument the smart money is having.
The future will regard this a picture of fraud, not a flywheel.Michael Burry, investor, on the AI industry’s cross-investments
The signals behind the AI noise, once a week
We track the funding, deals, and tech actually reshaping construction, minus the hype.
Join 3000+ ReadersWhat this means if you build or back construction tech
Cheaper AI helps your margins. A wobble hurts your funding.
For tech buyers, the cost crash is mostly good. The frontier models under pressure are literally in your stack. Bluebeam’s new Max tier runs on Anthropic’s Claude inside Revu. As inference gets cheaper, the AI features in tools you already pay for cost vendors less to run, which leaves room to pass savings on or ship more. If you’re evaluating AI tools, it’s worth asking vendors which models sit underneath and how exposed their pricing is to a single lab.
For founders and investors, the picture is sharper. The capital funding the AI boom overlaps with the capital funding ConTech, so if the AI trade cools, later-stage rounds and valuations tend to feel it first. There’s a public-opinion drag building too: a Gallup poll found 71% of Americans oppose an AI data center near them, higher than opposition to a nuclear plant, which is already slowing the buildout the whole model economy depends on.
None of this changes the real bottleneck in construction, which is adoption, not model price. As we covered in the quiet brake on ConTech AI, buyers move at the speed of procurement and change management, not the speed of the model. Cheaper AI doesn’t fix that. It just lowers the cost of waiting, which is quietly one of the more useful things to happen to this market in a while.
| What to compare | US frontier models (Claude, GPT) | Chinese open-weight models (GLM, Kimi, DeepSeek) |
|---|---|---|
| Price | Premium; around $5 / $25 per million tokens | Roughly 5x cheaper; around $1.40 / $4.40 per million tokens |
| Everyday coding performance | Top tier | Within a few points on standard benchmarks |
| How you run it | Vendor API | Open weights you can self-host and keep in-house |
| Provenance | The originators | Accused of distilling US models, which they deny |
| Availability | Can be hit by export-control shutoffs | Freely downloadable, harder to restrict |
| Who’s leaning in | Default for most US enterprises | Coinbase, Lindy, and a growing list rerouting work |
Close to it. When DeepSeek’s low-cost model landed in January 2025, roughly $1 trillion was wiped off US markets in a single trading day, with Nvidia alone losing about $589 billion, the biggest one-day drop for any company in US history. Nvidia recovered most of that within weeks, so it showed how fragile confidence in the AI trade is rather than destroying $1 trillion for good. (Source)
Open-weight models, mostly from Chinese labs, have caught up enough on everyday tasks to compete at roughly a fifth of the price, scoring within a few points of top US models on standard benchmarks. Because the weights are downloadable, companies can route routine work to them and self-host. The frontier labs still lead at the very top end, which is why firms keep the pricier models for the hardest jobs. (Source)
Distillation means training a smaller model on the outputs of a bigger one. It’s routine when a lab does it to its own model. The fight is about doing it to a competitor’s model without permission. Anthropic has accused several Chinese labs, and then Alibaba, of running tens of millions of exchanges through fake accounts to extract Claude’s capabilities. The accused firms deny it. (Source)
Parts of the worry are solid, parts are argued. An MIT study found 95% of enterprise AI pilots showed no measurable profit impact, and Oracle just had its worst week since 2001 over AI-related debt. The circular-financing concern raised by Michael Burry is real but contested by Nvidia and others. The demand and the debt are both real, the open question is timing. (Source)
It depends on how you run them and what data they touch. Self-hosting open weights keeps your data in-house, which is different from sending it to a foreign API. But provenance and licensing questions remain, and regulated sectors face extra scrutiny. It’s a procurement question with real teeth, not a simple yes or no. (Source)
Lower running costs. Many construction tools now embed frontier models, so cheaper inference means those features cost vendors less to run, which can flow through to pricing or more capability. It does not solve adoption, though, which is still the real ceiling in construction. (Source)
It’s worth watching. The capital funding the AI boom overlaps with the capital funding ConTech, and public opposition to the buildout is real, with 71% of Americans opposed to an AI data center near them. If the AI trade cools, later-stage ConTech rounds are the most likely to feel it first. (Source)
Related Articles
“You Can Only Go As Fast As Your Customers”: The Quiet Brake on ConTech AI
Bluebeam Max Goes Live: AI Lands Inside the Construction Workflow
CNBC: Nvidia’s record single-day loss on DeepSeek
Anthropic: Detecting and preventing distillation attacks
CNBC: Anthropic accuses Alibaba of illicit AI extraction
The New Stack: Coinbase halves its AI bill
Fortune: MIT report on enterprise AI pilots
CNBC: Oracle’s worst week since 2001
Gallup: Americans oppose AI data centers in their area
