The AI price war has begun

THE debate over open and closed artificial intelligence models is often dressed up as a philosophical contest. In reality, it is becoming a price war.

Chinese developers, such as DeepSeek, Alibaba’s Qwen and Moonshot’s Kimi, are releasing increasingly capable open-weight models at a fraction of the cost charged by American frontier laboratories.

“Open-weight” means users can download the trained parameters and modify them, subject to licensing terms. It does not imply access to the training code, detailed data information or complete recipe.

For customers, that distinction may matter less than the bill.

An open-weight model can be hosted privately and customised without paying the developer for every token consumed. It does not need to outperform the best model on every benchmark, it only needs to be good enough for the job.

That is already happening. Amazon Web Services offers Chinese models from DeepSeek, Qwen, Kimi, GLM and MiniMax through Bedrock. These are no longer obscure alternatives lurking in a developer’s basement. They are being distributed through the world’s largest cloud platforms.

America has already blinked

This puts OpenAI, Anthropic and Google in an awkward position. They must spend heavily to remain at the frontier while competitors steadily reduce what customers will pay.

The response will not be to give away every crown jewel.

The leading American models will remain closed and expensive, with the premium justified by better reasoning, reliability, security and integration. Alongside them will sit cheaper and sometimes open-weight models for workloads where “pretty good” is perfectly adequate.

America has already blinked.

OpenAI released its GPT-OSS open-weight models in 2025. Google has Gemma, while Meta has long used open distribution to challenge rivals with stronger cloud businesses.

The question is no longer whether US laboratories will open some models. It is how much capability they can release without undermining their own profitability.

They have little choice.

AI becomes expensive once companies move beyond cheerful chatbots and into agents.

These systems call models repeatedly, inhale lengthy documents and ask one model to check another’s work. The industry calls this “tokenmaxxing”, a glamorous name for leaving the meter running.

Until recently, the answer to weak performance was more context, more reasoning and more agents. That era may be ending.

In June 2026, researchers introduced latent context language models, which compress long inputs by four to 16 times.

Agents can skim the compressed material and expand only relevant sections. The idea is simple: Stop sending the entire filing cabinet when the model needs only one document.

Companies will eventually care less about benchmark scores and more about the cost of completing a useful task. Cheaper models and better context management can sustain adoption, while the industry searches for returns on the billions already committed.

Wall Street joins the party

Those returns cannot remain theoretical forever.

American technology companies may tolerate heavy investment longer than most industries, but they are still capitalist enterprises. Investors are already asking whether AI revenues justify the cost of chips, data centres and electricity.

Nvidia is reportedly working with Wall Street companies to mobilise more than US$500 billion for AI infrastructure. This may reduce the risk on the balance sheet, but it does not make that risk disappear.

It merely forwards that risk to private credit funds, insurers, banks and infrastructure investors.

The uncomfortable question is what the equipment will be worth, or its residual value, if AI usage stumbles. Think of residual value as the resale price of a car at the end of its loan. Lenders will refinance an asset if they know it can be resold at a reasonable price.

Graphics processing units are trickier.

A new chip or more efficient model could render older hardware obsolete much faster than expected. Wall Street may have eased the financing problem, but it has not eliminated the risk. Silicon could age less gracefully than a Japanese sport utility vehicle.

While Wall Street’s involvement has eased the immediate financing constraint, it has not repealed the laws of return on capital.

China faces the same problem from the opposite direction. Its developers have sacrificed profits to gain adoption and market share. That works when the objective is to build influence across the technology stack, rather than maximise the margin on every model.

However, DeepSeek’s recent price increases are revealing. They do not mean the price war is over as its models remain competitively priced. But even digital revolutionaries eventually need to generate cash flow. Chinese developers also need a sensible return on investment.

A ban without calling it a ban

Geopolitics will prevent this contest from producing one global winner. Some American companies will use Chinese models because the savings are compelling. Many others will not touch them, even when those models are hosted on US infrastructure.

A blanket US ban on private-sector use is possible, but it is not my base case.

Washington can achieve much of the same result through procurement rules, security requirements and restrictions covering defence, critical infrastructure and sensitive data. Congressional scrutiny of Chinese open-weight models is already increasing.

For many boards, the decision will be simpler. The savings will not justify having to explain the use of a Chinese model in the aftermath of a data breach. The reputational risk remains too great, not to mention the career implications for the chief information officer.

The model may not be the winner

The foundation-model market will probably consolidate around two or three large players across the US and China. That does not mean the rest of the AI market will be winner-takes-all.

There will be category winners where specialised data, interfaces and workflows matter more than owning the largest model. Wispr and ElevenLabs in voice, Gamma in presentations, Suno in music and Notion in enterprise productivity illustrate the pattern.

Foundation models may supply the intelligence, but these applications turn it into products people want to use. They do not need to win the model race. They need to win their niche.

Cloud platforms may also be among the safer winners.

Amazon, for example, can offer competing models and collect a toll regardless of which one customers choose. Application companies that can switch easily between models should also benefit as intelligence becomes cheaper.

The unresolved issue is responsibility. A closed-model provider can monitor usage, update safeguards and revoke access.

Once model weights are released, that control is largely lost. Determined users can fine-tune open-weight models to bypass safeguards, leaving deployers to provide protections that are typically built into hosted services.

Open models will set the industry’s price floor. Closed models will try to justify a premium through capability, trust and convenience. Both sides can produce winners.

The real battle is not over who builds the cleverest model. It is over who can turn cheaper intelligence into durable cash flow.

The writer is head of investment strategy, UOB Private Bank