Tencent shipped a near-frontier AI model on Friday, four months ahead of Bernstein’s schedule. Hy4-preview more than doubles its predecessor’s parameter count; in Citi’s characterization, it posts benchmark rankings on long-horizon task execution that match or exceed Chinese peers running larger models — Alibaba’s Qwen3.8-Max and Moonshot’s Kimi K3, both higher in total parameters. Bernstein, separately, calls it “probably the first Tencent model that’s a serious near-frontier release.” Hong Kong’s AI underdog joined the frontier conversation ahead of where the street had it. The equity tape flinched on Federal Reserve Chair Kevin Warsh’s Jackson Hole hawkishness the same morning. The structural fact underneath both reactions is that the consensus frame for sizing Asia’s AI capex just stopped fitting the data.

It is worth being precise about what “near-frontier” means here, because the public discourse has a habit of treating “AI model” as a thing rather than a specific architecture, trained on specific data, and evaluated on specific benchmarks. A model is frontier-grade when its capability band on a defined task category overlaps with the leading systems — Claude, GPT, Gemini at the top end — and is competitive without leading on any single benchmark. Bernstein’s phrase “serious near-frontier” is calibrated, not promotional: the model belongs in the conversation, not at its head. Citi’s read is narrower and more technical. Hy4-preview is smaller in total parameters than Qwen3.8-Max and Kimi K3. Its benchmark rankings on long-horizon task execution are comparable to or exceed those peers. That is the engineering fact the consensus missed.

The “total parameters” framing matters more than it looks, because the Chinese labs are almost universally using mixture-of-experts architectures — MoE, in the literature — in which only a fraction of the model’s parameters are active on any given token. It is true that total parameter counts are the industry-standard comparison metric; every benchmark leaderboard, every analyst note, every frontier-lab press release quotes them. The trouble is that they conceal the active compute per inference, which is what determines cost. A MoE model with 750 billion total parameters might run closer to a third of that on any given token. Hy4-preview more than doubles its predecessor’s total parameters. The right question for analysts is what fraction of those parameters are active per token and what shape the inference cost curve takes at scale. That question has not been answered in the published notes. It will determine whether Hy4-preview’s “smaller than Qwen3.8-Max” framing is the right unit of comparison or the wrong one.

The phrase “long-horizon task execution” is doing real work in the Citi note, and most readers will skim past it. The capability it names is the ability to plan and execute multi-step tasks — agentic workflows where the model maintains coherent state across many turns, calls tools, recovers from intermediate failures, and produces a result that satisfies a complex specification. The benchmarks that measure this band — SWE-bench for software engineering tasks, tau-bench for agentic customer-service workflows, GAIA for multi-modal reasoning, HLE-style reasoning evaluations — were, until recently, the property of the largest dense-parameter models. Smaller systems caught up over 2025. That Hy4-preview, running fewer total parameters than Qwen3.8-Max, posts competitive rankings here is the fact the consensus frame did not model. Long-horizon execution was thought to require scale; the data this month says it does not, or not as much of it.

That fact has a corollary the sell-side has not yet absorbed. If a model with fewer total parameters can match larger peers on long-horizon task execution, the inference compute required to deliver that capability is lower than the consensus assumes. Inference is where hyperscaler revenue actually lives: API calls, embedded integrations, internal productivity tools. A model that delivers frontier-band performance at lower token-cost shifts the unit economics of every product running on it. The Hy4 official release, per Citi, is “in coming months.” That timeline is when the next leg of this story lands, not Friday’s preview.

The supply chain underneath Friday’s preview has been preparing for it longer than the equity tape knows. ttb wealth securities raised its target price on Hana Microelectronics — a Thai electronics-manufacturing-services shop most foreign funds have never written down — to 55 baht from 48 on the news that two AI data-center orders are now in production and that two AI customers will sit in Hana’s top five by 2027. The share of sales attributable to AI orders went from a 7% prior assumption to 12% in 2027 and 13% in 2028. The shares trade at 47.50. That is the price action of a tier-2 EMS catching the overflow the hyperscalers cannot absorb at tier-1 capacity, and the qualification cycle that put those orders into production started twelve to eighteen months ago. The pipeline was committed long before Hy4-preview arrived. It will still be committed when the official version ships.

The same pattern shows up in the smartphone display market, which is the unglamorous sibling story to the AI capex boom and reads the same way. TrendForce now forecasts a 2.5% decline in global smartphone panel shipments to 2.25 billion units in 2026. Volume is soft. Mix is concentrating: BOE held 26.1% of Q2 market share, ahead of Samsung Display and TCL CSOT, with Apple and Samsung device demand holding the demand line. Consolidation at the top of a contracting market is the structural read; collapse is the tape read. BOE’s display capex is funding the next product cycle — foldable OLED for premium smartphones, larger-format panels for AI-enabled laptops — at the same time as AI server demand pulls tier-2 EMS overflow.

Which brings the Warsh tape back to where it belongs. SK Hynix fell 2.5% Monday. Samsung Electronics dropped 1.95%. TSMC lost 1.45%. It is true that equity multiples compress on higher discount rates, and that is what Monday’s tape priced. The trouble is that AI capex is funded by hyperscaler operating cash flow, not by equity market multiples. A 2.5% drop in SK Hynix reflects multiple compression; it does not reflect order-book weakness. The tier-2 EMS shops taking AI orders into production this quarter are not reading the same discount-rate tape. The capex plans underwritten by Hana’s new orders, BOE’s display build, and Tencent’s accelerated model release are the same plans underwriting the Warsh-shocked Korean memory names — viewed from different layers of the stack. The discount rate is one layer. The order book is another. Monday’s tape priced the first and ignored the second.

Two notes on what to watch. First, the Hy4 official release, expected in coming months per Citi, is the next data point that will force the sell-side to revise its China AI capability timeline. If the official version lands with the same benchmark profile as the preview, the consensus shifts by a quarter. Second, the Hana earnings print will show whether the two AI data-center customers move from “production started” to “production ramped.” ttb’s 12-13% sales assumption assumes the latter; the former is what the order book currently confirms. The deadline that matters is not Jackson Hole. It is the next earnings call.