Tencent did three things in one breath this month. It raised the free quota on Hunyuan Hy4. It extended Hy3's free allowance to November 15, 2026. And it expanded inference capacity tenfold β then told users to try again off-peak.
Read the sequence backwards. A tenfold expansion that still throttles is not a growth story. It is a confession about where the bottleneck lives. Peak concurrency broke the ceiling, and the ceiling was made of silicon, not software. For everyone who spent the last cycle arguing that compute is the new oil, here is the receipt: the oil is refined, the pipeline is thin, and the refinery belongs to somebody else.
The announcement carried six operational facts and zero technical ones. No parameter count. No context window. No benchmark placement. No architecture note. No word on whether Hy4 is open-weight or API-only. That absence is the most informative datum in the release. A launch that leads with quota math instead of capability math is a distribution play, not a research milestone. Hy3 staying in service β and staying free for another year β tells you the product matrix is layered, not replaced. Tencent is not shipping a frontier model and daring you to compare scores. It is shipping a funnel.
The funnel has an entrance, and the entrance is not the model. It is WorkBuddy, and behind WorkBuddy the WeChat and WeCom surfaces that no independent lab can rent at any price. The vertical play here β model plus office workflow β mirrors ByteDance's Doubao-and-Feishu pairing and Alibaba's Qwen-and-DingTalk pairing. Set that against the Chinese lab price war, where free tiers have become the default entry price for developer mindshare. Every quota extension is a subsidy with a timer on it. The timer here runs past 2026. That is not generosity. That is a burn rate dressed as a public good.
Here is the mechanical part, and it is the part crypto keeps getting wrong.
Inference is now the dominant compute load in production AI. Training is episodic and batchable. Inference is continuous and latency-bound. The two are not the same capacity-planning discipline. You can queue a training job for a week. You cannot queue a user's prompt for a week and keep the user. When Tencent writes "peak concurrency," it is describing a curve with a sharp head and a long tail β average utilization looks healthy, and the head still punches through the ceiling every evening at nine.
Tenfold expansion of what, exactly? Node count, GPU cards, or quota ceiling? The word "tenfold" is doing a great deal of unexamined work in that sentence. In my years of reading postmortems and vendor notices, a stated multiplier without a stated base is almost always a ratio of the flattering quantity. If the base was a conservative pilot allocation, ten times a pilot is still a pilot.
Then the honest admission: high-end compute is constrained. In the Chinese context that phrase maps onto advanced-node accelerator supply β export controls on one side, domestic substitution still climbing its yield curve on the other. This is not a Tencent-specific wound. It is the structural condition of every lab operating under the same import regime. The binding constraint on AI capability in 2026 is not algorithmic. It is the number of chips you can physically rack, power, and cool.
Now the pivot that matters for this audience.
Every decentralized compute network I have reviewed over the past eighteen months sells the same slide: idle GPU supply, aggregated, permissionless, cheaper than hyperscalers. Some of that is real. Spot markets for inference do clear, and they clear well at average load. But look at what Hy4's throttle actually tests. Burst absorption. A global permissionless market is a beautiful instrument for smoothing a long tail and a poor instrument for absorbing a spike, because spikes require pre-positioned capacity that somebody paid for in advance and held idle. Nobody in a token-incentivized network wants to hold idle. Idle is negative carry. Idle is a governance forum thread about emissions.
So the decentralized pitch has to move. It cannot win on raw burst capacity against a firm that will eat a tenfold capex hit to defend a funnel. It has to win on something a hyperscaler structurally cannot offer.
That something is proof.
The missing primitive is not cheaper inference. It is verifiable inference β a cryptographic or attestation-backed receipt that a specific model, at a specific version, actually produced the output you received. Right now you take the API's word for it. You take the throttle's word for it. You take the silent model swap's word for it. During the 2017 ICO cycle I spent three months auditing contracts for a DAO that wanted to democratize venture capital, and I learned this lesson in a different register: reviewing code tells you what the contract permits, not what the operator does. Code as conscience only works when behavior is observable. Audit the algorithm, not just the code.
Apply that to inference. A provider can downgrade you to a smaller checkpoint at peak, route you to a quantized variant, or shift sampling parameters, and you would never know. There is no reentrancy bug to find because there is no state to inspect. The output is the only artifact, and the output is unauthenticated. Signed model manifests, TEE attestation, optimistic challenge windows β that is not a cheaper product. It is a different product. It is the only version of this market where "which model answered me" is a verifiable fact rather than a marketing claim.
I spent much of 2024 in rooms translating between protocol teams and institutional risk committees, and the question that came up in every single one was the same: how do we know what the model actually did? We reframed compliance there as transparent accountability rather than censorship, and it landed. The same reframing applies here. Attestation is not an ideological flourish. It is the minimum viable audit trail, and it is the one thing a vertically integrated funnel has no incentive to ship.
That is where value accrues. Not in the GPU. In the receipt.
Now the part that will annoy the believers.
Decentralized compute networks will, for the next several years, be worse than Tencent at precisely the thing Tencent just failed at. Burst capacity. Coordination across untrusted nodes costs latency, and latency is the product. A tenfold capex commitment is available to a balance sheet; it is not available to a token treasury without a death spiral. Anyone telling you otherwise is selling the same tenfold fiction, just with a governance vote attached.
And a second uncomfortable point. Throttling is a soft kill switch. When a provider decides who gets served at peak, it is making an allocation decision with political content β and the free tier is the most efficient mechanism ever devised for making that decision invisible. You are not censored. You are deprioritized. Same outcome, no headline. The sociology of the free tier is that it converts a governance question into a user-experience question, and users do not file governance complaints about latency. They file churn.
Trust no one, verify the solitude. The solitude here is the queue you sit in at nine o'clock and cannot see.
Watch three signals over the next two quarters. Whether Hy4 publishes a technical report or a benchmark placement. Whether the free tier converts to a priced tier before the 2026 deadline. And whether any inference provider, centralized or not, ships a signed and auditable model manifest.
The first two are business. The third decides whether AI remains a service you rent or becomes infrastructure you can check. Speed kills. Precision saves. And precision, in an inference market, means a receipt.