Short Bets on Chinese AI: A Market-Level Audit of the Price War

The ledger remembers what the interface forgets. On December 14, 2026, the market's short position against Zhipu AI and MiniMax reached an all-time high. The raw number — record short bets piling up — is not a headline. It is a data point. A signal embedded in market structure that tells a specific story about unit economics, competitive positioning, and the gap between narrative and cash flow. The source reporting comes from Crypto Briefing, a financial outlet with a crypto-native readership. That matters. It means the short interest data is being read through a speculative lens, not an operational one. But the underlying fact stands: institutional capital is betting against two of China's most prominent AI startups. This is not a technology story. It is a capital efficiency story wearing a technology costume. To understand what the shorts see, you have to understand the terrain. Zhipu AI builds the GLM series of large language models. MiniMax builds the MiniMax-01 architecture, focused on multimodal and long-context inference. Both are classified among China's 'AI tigers' — private companies valued in the tens of billions, funded by state-backed capital and global venture funds. Both compete in the API market against Baidu, Alibaba, and ByteDance, which can afford to run model inference at a loss indefinitely because their AI divisions are loss leaders for cloud and advertising revenue. The price war began in May 2025 when Alibaba cut Qwen API prices by 97 percent. ByteDance followed with aggressive volume discounts. Zhipu and MiniMax had no choice but to match. In a commodity market — and that is what LLM inference has become — pricing power is determined by who can sustain negative margins longest. The shorts are betting that Zhipu and MiniMax run out of cash before the giants blink. This is where my background becomes relevant. I spent six months auditing the Ethereum 2.0 Slasher protocol in 2017, and later dissected the MakerDAO CDP liquidation logic during the 2020 DeFi Summer. Both experiences taught me the same lesson: when a system's economic parameters are misaligned with its real cost structure, the failure is not a question of if, but when. The same forensic lens applies here. Let me break down the actual mechanics of the AI price war in operational terms. An API call on a mid-sized LLM costs roughly $0.30 per million input tokens at list price. After the price war, effective rates dropped to under $0.05 per million tokens for comparable quality. The cost to serve that request — GPU amortization, electricity, cooling, staff — is approximately $0.04 to $0.06 per million tokens for a company with efficient infrastructure. That means gross margins on API revenue are now hovering between negative ten percent and positive five percent, depending on utilization rates. For Zhipu and MiniMax, which do not own their own data centers and rely on leased compute from Alibaba Cloud and Tencent Cloud, the cost structure is worse. They are paying retail for compute while selling at wholesale prices. Here is the core finding: the price war is not about model quality. It is about capital allocation. In the DeFi ecosystem, we saw the same pattern in 2021 when liquidity mining rewards created a race to zero. Protocols competed for total value locked by paying out unsustainable token emissions. The ones that survived were not the ones with the best technology — they were the ones with the lowest operational burn rate. Curve survived because its veTokenomics aligned incentives. Olympus collapsed because its rebase mechanism was a Ponzi structure disguised as a protocol. Zhipu and MiniMax are in the Curve position, but with one critical difference: they do not have a native token to inflate. They must burn real dollars. The short sellers know this. Their thesis is straightforward: Zhipu and MiniMax cannot outspend Alibaba or ByteDance on a per-customer basis, and they cannot out-innovate fast enough to escape the commodity trap. The shorts are not betting on a technology failure. They are betting on a treasury failure. This is the same logic that drove short positions against Three Arrows Capital in 2022. I spent three months tracing their liquidation cascades through Anchor Protocol and Venus Market. The insolvency was not a protocol bug. It was a leverage management failure. The same forensic pattern applies here. But here is the contrarian angle, and it is one the shorts may be missing. The price war has a second-order effect that is not reflected in the short thesis: it accelerates downstream adoption. When API costs drop by 90 percent, the volume of AI applications explodes. Every SaaS product, every content generation tool, every customer support bot becomes economically viable. The demand curve for inference is elastic. Lower prices mean exponentially more calls. Zhipu and MiniMax, despite their margin compression, are seeing record API utilization. The question is whether volume growth outpaces margin decline. In my audit of the OpenSea Seaport migration in late 2021, I identified a race condition in the consideration fulfillment logic that could have allowed front-running attacks on rare asset sales. The lesson was that infrastructure stability matters more than feature velocity. The same principle applies here. Zhipu and MiniMax may be losing money on every call, but they are building the infrastructure layer that applications depend on. The question is whether that dependency translates into durable revenue. There is also a second blind spot in the short thesis: government support. Zhipu AI is backed by Beijing municipal state capital. MiniMax has received funding from Tencent and Alibaba itself — the very competitors it is fighting. This is not a clean competitive battlefield. In China, the state has a direct interest in maintaining multiple independent AI labs. The shorts are treating this as a pure market competition. It is not. It is a managed oligopoly where the government has the power to intervene, subsidize, or consolidate at will. In 2022, when the Chinese government ordered the consolidation of ride-hailing companies, Didi's competitors were forced to merge. The same dynamic could play out in AI, with Zhipu or MiniMax being acquired by a state-backed entity at a premium. The market has not priced in this optionality. Short interest is a snapshot of current sentiment, not a forecast of structural intervention. If Beijing steps in with a national AI subsidy program or orchestrates a strategic merger, the short thesis collapses. Let me also address the infrastructure dependence issue, which the source article does not mention at all. Zhipu and MiniMax rely on NVIDIA GPUs for training and inference. Export controls have made H100 and A100 acquisition difficult and expensive. Both companies are actively adapting their models to run on Huawei Ascend 910B chips. If that adaptation succeeds — and my assessment of the technical difficulty is moderate — their cost structure improves significantly. Domestic chip prices are 40 percent lower than imported GPUs. This is a swing factor that could fundamentally alter the unit economics of their API business. The shorts are not accounting for this. In my work defining the AI Agent Payment Layer Specification in 2026, I insisted on backward-compatible design using proven cryptographic primitives rather than experimental tokenomics. The principle was: do not bet on unproven technology when established infrastructure can do the job. The same principle applies to Chinese AI infrastructure. Huawei's Ascend chips are not as performant as NVIDIA's, but they are available, and they are cheap. For inference workloads — which is where Zhipu and MiniMax generate the bulk of their volume — the performance gap is narrowing. The cost advantage could give them a pricing edge that the shorts have not modeled. Here is my assessment, stated plainly. The short thesis is correct on the timeline of the next six to twelve months. Zhipu and MiniMax will continue to face margin compression. Their API revenue growth will not offset the cost of the price war. They will need to raise capital at down rounds or accept strategic investments from competitors. This is the base case. The alternative case — and this is where I diverge from the market consensus — is that the price war forces a consolidation that benefits the survivors. The companies that can survive two more quarters of negative margins will emerge with dominant market share. The Chinese AI market is large enough to support two or three independent players. Zhipu has the government relationship. MiniMax has the product distribution through its consumer-facing apps. Neither is a guaranteed loser. The real signal to watch is not the short interest. It is the funding pipeline. If Zhipu or MiniMax announces a new round at a flat or down valuation within the next 90 days, the shorts are validated. If they announce strategic partnerships with cloud providers that include compute credits, the calculus changes. The ledger remembers what the interface forgets: capital is the ultimate audit trail. The market is pricing in a specific outcome. The question is whether the market's model of the world is complete. My experience auditing the Slasher protocol taught me that consensus failures are rarely caused by a single bug. They are caused by cascading assumptions that were never validated under stress. The same is true here. The shorts have identified a real vulnerability: cash burn. But they have not modeled the full state space. Government intervention, domestic chip adoption, and downstream demand elasticity are all variables that could invalidate their position. The market is never fully efficient. It is just insufficiently lazy. The shorts have done their homework on the balance sheet. They have not done their homework on the policy layer or the hardware supply chain. That asymmetry is where the opportunity lies. Here is the takeaway. Treat the short position as a stress test, not a verdict. The next ninety days will determine whether Zhipu and MiniMax are Curve or Olympus. The difference between survival and collapse is not model quality. It is capital discipline, government support, and the ability to adapt to a hostile cost environment. The market has made its bet. The ledger will tell us who was right.