Kimi K3, the latest flagship model out of Chinese lab Moonshot AI, has been out for a few weeks and has already produced a claim doing the rounds online: that a model this capable, priced this competitively, breaks the economic case for AI altogether, because frontier performance no longer requires a frontier bill. It would be a convenient story if it were true. The research suggests the opposite conclusion, and the direction of travel matters more to your AI budget next year than the headline price does.
I wrote in July, when DeepSeek cut its API prices 75% in one move, that falling model prices are a genuine reason to run more small experiments, not a reason to expect AI adoption itself to get cheaper. Kimi K3 sharpens that argument rather than overturning it, and it adds a second, less obvious wrinkle: even a model whose weights are given away free is not actually free to run, and someone is paying for the difference.
Cheaper tokens don't mean a cheaper bill
Kimi K3 has taken the lead on independent coding benchmarks, the first time a Chinese model has done so, and Microsoft engineers are reportedly testing whether it can be dropped into Copilot. Analysts at Exponential View, publishing research this year on what they call the state of the AI economy, found that token usage is elastic: every 10% cut in the price per token produces something like a 12–18% rise in the volume of tokens consumed, not the other way round. A separate academic study (Demirer et al., cited in the same research) found a near-identical effect, an 11% rise in volume for every 10% price cut. Put plainly, cheaper AI does not reduce what organisations spend on AI. It increases it, because people and systems use more of a thing once it costs less, and the rise in usage outpaces the fall in price.
Stop budgeting for AI as a fixed monthly line item that will shrink as vendors cut prices. Budget instead for usage to grow faster than the price falls, and build a usage cap or monthly review into any AI tool you roll out, the same discipline you'd apply to a cloud-hosting bill, so a "cheaper" model doesn't quietly become a larger invoice six months from now.
"Open-weight" is not the same as "free to run"
The second claim worth testing is that because Kimi K3's model weights are open and downloadable, running it costs next to nothing. Kimi K3 has 2.8 trillion parameters; the weights alone occupy 1.4 terabytes, and serving the model at any useful scale needs something like a rack of 72 high-end GPUs, which costs $3–4 million to buy and install and consumes roughly 120 kilowatts continuously, over a million kilowatt-hours a year. Renting the equivalent capacity on the open market would run to roughly $7 million a year, before networking, storage, cooling and the people to run it. The model is free in the sense that nobody is charging you a licence fee. It is not free in the sense that matters to your invoice: somebody, an infrastructure provider or an API reseller, is paying for that electricity and hardware, and it finds its way into what you're quoted per token, just without a line item that says so.
A potential client and I had exactly this discussion a few weeks ago: what does using AI actually cost the company, once you get past the number on the pricing page? We worked through the options together, running something locally, individual subscriptions to a frontier model, an enterprise tier, and it became obvious fairly quickly that without in-house AI and hardware expertise, some of those options simply weren't viable for them, especially with the landscape changing as fast as it currently is. The same tension came up again when we set frontier models against open-weight ones. Whichever side of that line you're on, the cost that doesn't go away is compute. SMEs need the right information to make an informed choice here, and vendors need to be honest about the pros and cons of each option, because none of them, frankly, is perfect.
Why this matters for Malaysian SMEs
Almost no Malaysian SME runs its own inference; you're buying access to a hosted model through a vendor, which means the question that matters is not whether the underlying model is open or closed but whether the price you're quoted reflects genuine cost or a promotional rate struck during a price war between the major labs and Chinese contenders like Moonshot and Alibaba (whose 2.4-trillion-parameter Qwen3.8 was announced days after Kimi K3, with no pricing yet attached). In assessment conversations, budget planning almost always starts from today's per-token price and assumes it can only fall from here. The more useful assumption, on this evidence, is that your total spend will rise as usage grows even while the unit price keeps falling, which changes what you negotiate for in a vendor contract: ask for usage tiers and cost alerts, not just a low headline rate that can change or get subsidised away once a price war cools.
Want the next one of these in your inbox?
Subscribe on the Insights page for practical perspectives on AI in Malaysia, no hype, no spam.
What This Means for You
Kimi K3 is a genuinely impressive, genuinely cheap model, and prices across the industry will keep falling, more or less as they have for two years now. Treat every "AI just got cheaper" headline as an invitation to plan for more usage rather than less spend, and treat "open-weight" as a licensing detail rather than a cost guarantee, because the electricity bill for running 2.8 trillion parameters doesn't disappear just because nobody's charging you for the weights.