On August 26, Reuters reported that Moonshot AI, the Beijing startup behind the Kimi models, is in early-stage talks with Microsoft, Amazon and Alphabet’s Google over hosting its flagship open-weight model, Kimi K3, on Azure, AWS and Google Cloud. The price of admission: Moonshot wants up to 30% of the revenue those platforms generate from K3-related services.

Nothing is signed, all four companies declined to comment, and the unresolved items are the predictable ones — how the revenue is defined, how much data access Moonshot gets, and how token usage gets audited.

Still, the number is not the story. For two years the money in this industry has flowed one way: model companies pay clouds. They buy compute; they sign investment agreements that route value to their backers; OpenAI reportedly hands Microsoft around 20% of its revenue under an arrangement Microsoft disclosed in April that runs to 2030. A Chinese lab asking the three largest American clouds to pay it an upstream royalty inverts that flow. If it lands, it would be the first large, explicit revenue-sharing deal between a Chinese AI company and the big three — not a model quietly appearing in a marketplace, but a negotiated seat in the value chain.

What Moonshot is actually selling

Start with what K3 is not: scarce in the way closed models are. It is open-weight — 2.8 trillion parameters anyone can download — and while purists note it does not fully meet the definition of open source, the distinction matters less than the economics. Weights that can be copied cannot be sold at a premium. What Moonshot sells instead is everything around the weights: the officially maintained version, enterprise support, service-level agreements, compliance, a single bill. And the model’s own scale does the selling. Almost no enterprise can deploy a 2.8-trillion-parameter model efficiently on its own hardware; the clouds can.

The rest of the leverage is performance, and it is perishable. Third-party evaluators have rated K3 comparable to OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8, especially on multi-step tasks, and the model has drawn international attention since its July release. But leaderboard leads in this industry last months, not years — which is precisely why Moonshot is negotiating now, while K3 is still a reason for a cloud to pick up the phone.

The arithmetic of open weights

Open-weight monetization has always leaked. Three holes, in rough order of damage. Distribution cannot be locked: anyone who downloads the weights can host the model and compete with you. Price cannot hold: DeepSeek-style discounting has dragged per-token pricing across the industry toward cost. And the customer is not yours: relationships sit with clouds and aggregators like OpenRouter, which know exactly who is calling the API while the lab often knows only how often.

Channel deals themselves are not new — Zhipu, Mistral and Meta have all arranged cloud hosting before, and Moonshot has reportedly already signed smaller versions of such deals with minor platforms. What is new in the Kimi approach is the shape of the ask: one unified, public, upstream cut, negotiated simultaneously with all three American hyperscalers. The structure is deliberately tiered. Small developers keep the free download; the bill arrives only when a platform’s or customer’s volume crosses a threshold. Distribution and monetization, decoupled.

The clouds can afford the rent. One research note, citing SemiAnalysis cost modeling, puts K3’s blended realized price at about $0.74 per million tokens against a serving cost of roughly $0.17. Run a dollar of K3 revenue through that math and you get about 30 cents to Moonshot, 23 cents of inference cost, and 47 cents of gross margin left for the cloud. Thirty percent is aggressive, not absurd — provided the cost model survives contact with real production traffic.

Willing to shelve is not willing to sell

Why would the hyperscalers play at all? A cheap, strong open model adds inference consumption and disciplines the pricing of the models they would rather sell you. But each of the three has core assets to protect. Amazon has committed tens of billions to OpenAI and Anthropic and is the exclusive third-party cloud distributor of OpenAI Frontier. Microsoft is OpenAI’s largest backer. Google has the most direct conflict of all: Gemini sits at the center of its cloud, Workspace and agent stack, and Google simultaneously supplies TPUs and Vertex distribution to Anthropic. For all three, K3 is shelf inventory, not a flagship.

Which is why listing and distributing are different verbs. The clouds own the bill, the customer relationship, the shelf ordering and the termination clause. Moonshot’s revenue will be decided not by whether K3 appears in a catalog but by whether it gets pulled into enterprise sales motions and agent product stacks — none of which the clouds are obliged to grant.

The borrowed boat

Chinese models have been abroad for a while, in layers. The first layer is weights loose in the developer community. The second is listings on OpenRouter and cloud marketplaces. Only the third — formal hosted revenue shares, co-selling, enterprise billing — is where money actually changes hands. K3 is attempting the jump from the second layer to the third, and the reason it needs help is structural. Moonshot has no global salesforce, no rails for collecting dollars, no overseas compliance footprint. Inside a cloud contract, by contrast, a customer ticks a box on a budget that procurement approved long ago.

Hence the paradox the deal crystallizes: open weights bought Chinese models their global influence, but monetization still has to run on American rails — hosted by American companies, billed through American contracts, on routes regulated in Washington. The boat, the docks and the settlement system all belong to someone else. The near-term economics are correspondingly modest: the same note’s base case has initial revenue in the $10–20 million range against an ARR around $300 million — under a tenth of it. Crossing $100 million requires on the order of 500 trillion tokens a year. The real value of a signed deal is validation, and the template it sets: if Moonshot collects, Qwen, GLM and DeepSeek have every reason to reopen their own negotiations, and the competition between Chinese labs quietly shifts from download counts to revenue attribution.

Washington holds the tiller

The risks are ranked, and commercial terms are not first among them. Treasury Secretary Scott Bessent said last month he could consider adding Moonshot to a trade blacklist; U. S. officials have accused the company of distilling K3 from Anthropic’s Fable model and of illegally obtaining Nvidia chips, charges Moonshot rejects. A revenue-split percentage does not survive a blocked transaction. Then the definitional risk: 30% of what — gross revenue including compute, or the model’s net take after channel fees? A signed contract could still land lower, floored, or regionally scoped. Then the window: Qwen, GLM and DeepSeek’s next versions are close behind, and scarcity is the only thing that makes a royalty rational. Finally, signing is not collecting: if no major cloud closes within three to six months, or K3 never enters co-selling, the thesis quietly downgrades.

The research note behind these numbers calls the deal paradigm value over earnings value, and that is right — with one precision. The paradigm being tested is not “Chinese models go global.” The weights already went global; weights don’t need a visa. The paradigm is narrower and stranger: whether a Chinese model can become an upstream royalty earner on American infrastructure. Kimi is not negotiating for access — access it already grants away, free, forever, to anyone who clicks download. It is trying to convert adoption into property. The ship is borrowed, the harbor is American, and whether the freight ever gets paid will be decided in Washington before it is decided on any benchmark.

I watched it so you don’t have to sit through all five hours (though if you read Chinese, it’s worth it). Here’s what stayed with me.

Who is Liao Heng?

His biography reads like a compressed history of the global chip industry itself. Liao entered Tsinghua University’s computer science program at 14 as part of its gifted youth class, did postdoctoral work at Princeton, and began his chip career in 1997 at PMC-Sierra, the Canadian semiconductor company. In 2016, he joined Huawei’s HiSilicon, where he led the launch and development of the Ascend AI chip project — making him one of its founding architects.

He insists, repeatedly, that what follows is “just an engineer’s story.” It is not. It is a story about what happens when an entire supply chain is weaponized against you — and what you build instead.

The 18-story pagoda

The interview’s most durable contribution is a mental model Liao calls the “18-story pagoda” (18层宝塔) of the semiconductor industry.

Picture the chip industry as a tower. The basement and lower floors are the foundations: materials, manufacturing equipment, EDA software, process technology. The middle floors are the chips themselves — the connective layer. The upper floors are systems, software, algorithms, and the end products consumers touch.

The point of the model is that no single floor wins alone. American dominance, in Liao’s telling, was built by controlling the foundations (ASML, Applied Materials, Synopsys, TSMC’s process leadership). But the profits — and increasingly, the strategic leverage — concentrate at the top, where Nvidia doesn’t just sell chips but an entire software-and-system stack. His warning: when application-layer players monopolize value, they squeeze the margins of everyone below them, and compute competition becomes a contest between entire stacks, not individual chips.

This framing explains Huawei’s whole post-2020 strategy. If you can’t match your rival floor-by-floor at the foundation — and after the export controls, Huawei couldn’t — you compensate through cross-layer coordination: make the floors work together better than your competitor’s do.

Engineering your way around physics

The most technically striking part of the interview is Liao’s account of how Huawei answered a brutal question: what do you do when you’re locked out of the most advanced process nodes?

His answer, in essence: stop chasing single-chip performance and treat the system as the unit of optimization. Four pillars:

1. Ultra-high-bandwidth interconnect. Instead of treating chip-to-chip links as plumbing, Huawei made interconnect a first-class design metric. Liao’s formula is almost disarmingly simple: system compute = single-chip performance × cluster scale at which you can actually sustain it. If your chip is weaker but you can wire ten thousand of them together with less loss than anyone else, you can still win the room.

2. “Logic folding” advanced packaging. Using 2-micron hybrid bonding, Huawei stacks dies vertically with on the order of ten million connections between them — versus the tens of thousands typical in mainstream packaging — effectively folding what would be a flat chip into a dense 3D structure.

3. Hardware-software co-design. Huawei’s CANN software stack is adapted at the lowest levels to domestic large models, pushing hardware utilization (MFU) far higher than a generic stack would manage on the same silicon.

4. Cluster-level design. Supernode architecture, cooling, power delivery, and scheduling designed as one organism — the “SuperPoD” approach now visible in the Ascend 950 generation.

Liao also offered a quietly radical take on Moore’s Law: by the 7nm node, he argues, its economic returns were already nearing exhaustion. The industry, he suggests, is shifting from compressing space (smaller transistors) to optimizing time (latency, utilization, scheduling) — a game in which system architects, not just fabs, hold the cards.

The darkest moment, and what came after

Liao is candid about the human side. When he joined Huawei, he admits, he carried what he calls the arrogance common among engineers trained abroad — a reflexive belief that certain hard problems could only be solved “over there.” He also admits to three significant misjudgments after joining.

What changed his mind was watching the organization execute under siege. After 2020, when TSMC could no longer manufacture Huawei’s designs and Ascend’s future looked terminal, the team regrouped — and today, Liao says, the group tapes out on the order of 70 new chips a year with virtually none coming back with fatal flaws, a success rate he describes as extreme by global industry standards.

There’s a pragmatic streak running through his assessment of where China now stands. In a decoupled world, he suggests, China would not run out of compute — its chips would simply consume more power. And with electricity prices roughly a quarter (or less) of those in other major economies, that trade-off is survivable. It’s a notably unsentimental way to think about technological sovereignty: not parity, but adequacy plus compensating advantages.

The talent anxiety

For all the geopolitics, the moment Liao sounds most genuinely worried is when the topic turns to students. He gave three reasons for breaking HiSilicon’s media silence: to leave some lessons from an engineer’s story; to reflect on the industry as a whole; and — the one he lingered on — because he fears a coming talent gap in hardware.

“It’s become very hard to get students interested in computer processor hardware,” he said. “On one hand I’m extremely surprised; on the other, I feel real anxiety. If nobody wants to learn this anymore, will we lose the next generation entering the field? There could be a break in the pipeline of people.”

It’s a striking confession from someone sitting atop one of the world’s largest chip design operations: the scarcest resource isn’t lithography machines or HBM. It’s young people willing to spend a decade learning how processors actually work.

The DeepSeek endorsement

One passage ricocheted around Chinese tech media: Liao’s praise for Liang Wenfeng, founder of DeepSeek. While the rest of the industry treated Scaling Laws as scripture — pile on parameters, pile on GPUs, trust that size crushes everything — Liang chose sparse-activation architectures and focused on high-value compute, anticipating the compute bottleneck before it arrived.

Liao’s framing of why this matters is worth quoting in spirit: Liang wasn’t doing algorithmic fine-tuning. He made a top-level strategic choice spanning software and hardware, and then did the painful work of making a harder-to-converge approach actually train. “He had foresight — he didn’t wait until five years later when compute ran out and there was no road left. He solved, in advance, a problem he expected. That’s the value of a pioneer.” The interview also effectively confirms what industry watchers had suspected: DeepSeek has been working closely with domestic chipmakers on adaptation and optimization.

Why this interview matters

Strip away the specifics and the interview is an artifact of a bigger shift. For thirty years, the semiconductor story was told from the foundation layers up — whoever held the best process node held the narrative. Liao Heng is making a different claim, from the middle of the pagoda: that the era of single-point advantage is ending, and the era of system integration as strategy has begun. Nvidia understood this early with CUDA and NVLink; Huawei was forced into it by sanctions; DeepSeek arrived at it from the algorithm side.

Whether or not you buy every part of Huawei’s version of events, the map Liao draws is useful for anyone trying to understand where AI compute competition actually goes next. The chip war, in his telling, won’t be won at 2nm. It will be won in the wiring, the software, the power grid — and, most of all, in whether enough young engineers still want to learn how the tower is built.

In a recent expert interview, a senior Alibaba insider with a decade of experience in cloud computing and GPU R&D offered a rare, detailed look at how the company is approaching AI chip procurement in 2026 — and what it reveals about the broader competitive landscape of domestic Chinese silicon.

A Capex Plan Measured in Hundreds of Billions

Alibaba’s 2026 capital expenditure is expected to land between RMB 170 billion and RMB 180 billion, staying just under the RMB 200 billion mark. However, if the operating costs associated with subleased high-end Nvidia GPUs — such as the B-series — are included, the total would exceed RMB 200 billion. For 2027, the expert said it is too early to give a precise figure, but RMB 200 billion serves as a working baseline.

Within this envelope, Alibaba plans to purchase roughly 400,000 of its own self-developed AI accelerator cards in 2026, at a cost of more than RMB 70 billion. Notably, only about 120,000 to 150,000 of these new cards are earmarked for internal use, including large model training. The reason internal consumption is relatively modest is that Alibaba has been deploying its own chips since 2023, building up a large installed base with a roughly five-year lifecycle. Internal teams are also under strict cost discipline, squeezing maximum utilization from existing hardware — for example, using software-based sharing to split a single card across multiple inference tasks. The bulk of new card purchases, by contrast, flows into the public cloud business, which serves external customers and absorbs hundreds of thousands of additional cards each year as a matter of course.

The T-Head Portfolio: Strong at Inference, Catching Up at Scale

Alibaba’s in-house silicon strategy rests on clear functional segmentation. The company’s Hanguang 800 chip excels at inference: its performance is comparable to Nvidia’s T4 and, in certain optimized scenarios, can reach 1.2 times the T4’s throughput. Its training capability, however, is weak — roughly 60 to 70 percent of a T4. As a result, internal practice is to use cutting-edge rented hardware — Nvidia’s B-series or H200, accessed through compute-leasing arrangements — for large model pre-training, while newly purchased self-developed chips handle fine-tuning, refinement, and reinforcement learning workloads.

Figure: Relative performance of leading domestic AI chips versus Nvidia reference cards, per the expert's estimates. Bars show estimated performance ranges; the dashed line marks parity with the relevant Nvidia baseline (T4 or A100).

Looking ahead, T-Head’s newer 890 and 950 series place it in the first tier of domestic AI chips alongside Huawei, according to the expert. The two companies are currently the only players in the industry capable of building cards with more than 140GB of HBM memory. That distinction matters enormously: for large-scale distributed pre-training clusters, the two non-negotiable technical requirements are inter-chip interconnect bandwidth of at least 700G and sufficiently large HBM capacity. Once those thresholds are met, single-card performance becomes less critical, because a 10,000-card cluster running at 95 to 97 percent scaling efficiency can deliver excellent aggregate performance. On this technical basis, the expert ranks Cambricon and Hygon in the second tier, with newer entrants such as Iluvatar CoreX’s Tiangai 300, Biren’s BR200, and MetaX in a third tier.

Huawei and Hygon: Confirmed Purchases, but With Limits

Alibaba has two confirmed domestic chip suppliers beyond its own silicon: Huawei and Hygon. The driver is the company’s government and enterprise business — particularly private cloud appliance projects — which generates more than RMB 20 billion in annual revenue. Hardware accounts for roughly 55 to 60 percent of that figure, and GPUs dominate the hardware mix, translating into an annual GPU procurement budget of RMB 8 to 10 billion for this segment. Vendor selection here is largely path-dependent: customers who used Hygon chips (with their ROCm-based ecosystem) in the first phases of a project resist switching to T-Head’s CUDA-like ecosystem later, and vice versa for Huawei-based deployments.

Market rumors that Alibaba has purchased hundreds of thousands of Huawei chips are significantly overstated, the expert said — actual volumes are below 100,000 cards. The economics explain why: Huawei’s 950 PR model is priced at around RMB 230,000 to 240,000 per card, comparable to or even more expensive than Nvidia’s H200, while the DT model approaches RMB 300,000.

Absent a deep strategic partnership or guaranteed upstream capacity, a purchase of that magnitude would create enormous cost pressure if the cards could not be monetized. Large-scale procurement of third-party domestic chips, the expert argued, only becomes rational once a chip enters the public cloud system. Hygon’s CPU offers a template: after entering Alibaba’s public cloud, its volumes could plausibly grow from 30,000–50,000 units to 100,000 within two to three years. The private and hybrid cloud market alone, at RMB 20 to 30 billion in total size, cannot support six-figure GPU volumes — especially with Alibaba’s own chips competing for the same budget.

Procurement Is a Business Decision, Not a Patriotic One

Perhaps the most striking theme of the interview is that Alibaba’s chip buying is governed by commercial logic rather than a “buy domestic at all costs” mandate. Cloud product managers, who own revenue targets for GPU instance services, evaluate chips on three dimensions.

• First, cost: Huawei cards run roughly 1.3 to 1.5 times the cost of Alibaba’s internal chips, and for physical GPU passthrough instances — where the card itself is the dominant cost — that premium flows directly into pricing and customer demand.

• Second, ecosystem maturity: public cloud workloads span gaming, audio-video processing, office automation and more, so broad compatibility is essential.

• Third, after-sales support: cloud services demand 24/7 stability, and a cautionary tale cited in the interview involved another cloud provider (”T Company”) that deployed some 2,000 Huawei cards in its Shanghai Songjiang data center, only to see the service falter when customer problems proved unresolvable and Huawei, viewing the project as too small, declined to provide sustained on-site support.

This commercial framing also explains why ecosystem gaps — not just process nodes or HBM supply — remain the core constraint on domestic chip adoption. Code porting is a persistent friction point: customers expect vendors to migrate legacy workloads, and smaller chipmakers lack the teams to do it. AI coding tools can automate perhaps 30 to 40 percent of porting legacy Fortran scientific programs, but full automation remains out of reach.

Supply Chains: HBM, SMIC, and the Nvidia Hedge

On supply, T-Head’s position is manageable but not without strain. SMIC’s allocated capacity is sufficient for the chip unit itself, but may fall short of the cloud business’s needs — making expansion or third-party chip purchases a likely move in the second half of 2026 if SMIC cannot scale. On HBM, domestic suppliers have made real progress: HBM3E was tested in China in 2025 with acceptable results, and companies such as ChangXin Memory are expected to deepen cooperation with T-Head. Alibaba would accept domestic HBM3E even with somewhat inferior thermals and efficiency — particularly for inference workloads — and expects pricing below overseas suppliers given strategic partnerships.

Meanwhile, the Nvidia channel remains a critical hedge. Alibaba’s compute-leasing business, projected at US$3 to 4 billion in 2026 and weighted toward Southeast Asia and Europe, has moved entirely to B200 cards since late 2025; future procurement may shift toward the B300. But with H200 supply potentially disrupted after July 2026, domestic chips are becoming a mandatory option for every Chinese cloud provider — even those, like Alibaba, with silicon of their own.

The article is based on an expert interview with a senior Alibaba cloud and GPU R&D professional

First half: MI300 still carries the business

Through the first half of 2026, AMD’s shipments were still dominated by the older generations. The MI450, though formally launched in H1, exists only as test samples — no volume shipments yet.

The workhorse is the MI300, at roughly 110,000-plus units in H1, going mainly to Azure, OCI, and Meta. The MI355 began its ramp in the same period with roughly 60,000-plus units, but it isn’t the lead product yet — production only started climbing this year. The MI250 added nearly 20,000 units, almost all to HPC and legacy customers expanding older MI200-era deployments; the expert was blunt that MI250 is an HPC part with mediocre AI credentials, weak on operator coverage and newer model architectures.

One more H1 item worth flagging: AMD moved about 30,000-plus units of MI308 — the export-compliant, cut-down MI300 — into China in Q1, racing ahead of restrictions from both Washington and Beijing. Once China clamped down on H20 and H200 class parts, MI308 went to zero. AMD has since shifted essentially all MI308 capacity back to standard MI300 production. That’s a one-time revenue bump that won’t repeat.

The customer map: OCI is the anchor, Meta is the hole

The customer breakdown is where this call got interesting, because the aggregate numbers hide a very lopsided distribution.

Microsoft (Azure) remains the longest-standing AI partner, with about 8,700 servers deployed — roughly 70,000 MI300-series cards. Its MI355 exposure is deliberately small: a few thousand units, described as small-batch trials rather than committed scale.

Oracle (OCI) is the opposite story. OCI wrote MI355-and-beyond into its contract from the start, so it gets delivery priority on the new silicon. It has taken close to 50,000 MI355 units in H1 — by far the largest allocation — while holding only about 1,400 MI300s from an early small order. OCI is, right now, the anchor customer for everything new AMD ships.

Meta is the disappointment, and the single biggest reason the full-year number came down. Meta had planned to be as aggressive as OCI. Instead, its H1 purchases amounted to roughly 2,000 servers — some 16,000 to 17,000 MI300 and MI355 cards — essentially trial volume. The problem wasn’t the product. Meta simply didn’t have data center space ready for AMD: nearly all of its H1 facility capacity went to Nvidia, whose deliveries to Meta were concentrated in exactly this window. The expert stressed that Meta’s underlying demand hasn’t changed and that follow-on orders are being considered. The volume is deferred, not lost — but deferred is deferred.

The remainder is a long tail: HPC shops, research institutions, universities, and small customers in Japan and Europe buying modest quantities of MI320 and MI250 — around 1,000-plus servers, or 10,000-plus cards, mostly for scientific computing.

Second half: the MI455 ramp, in ratios

H1 totaled about 200,000 units. The plan for Q3 and Q4 combined is close to 400,000, putting the full year above 560,000. Capacity is locked — AMD fixed its full-year 2026 allocation with TSMC in advance, and with its CPU lines running tight, there’s no slack to add more.

The composition of that second half is worth watching in ratios rather than absolutes:

• Q3 belongs to the MI355X, at roughly 100,000 units. The MI455 enters only as early samples and a first trickle of Helios racks — 120-odd cabinets at most, translating to under 10,000 chips. Call it 10:1 in favor of the 355X.

• Q4 is planned at roughly 6:4 — about 100,000 MI355X against 110,000 to 120,000 MI455 as the ramp takes hold. And there’s upside: if Q3 Helios feedback from early customers comes back clean, Q4 could shift toward 1:1.

That conditional is the whole game. The 455X’s Q4 volume is hostage to how the first hundred-odd Helios racks perform in the field this quarter.

On allocation, the pecking order for the 455 is clear: OCI first, Meta second, Azure third. Meta had originally signed for the 455 on roughly the same timeline as OCI; its facility delays pushed it down the queue.

Why did the year get cut from ~800K to ~560K? Three reasons, in order: Meta’s deferral; softer-than-expected overseas uptake, with customers in Europe, the Middle East, and Asia sitting in wait-and-see mode after lukewarm early-product feedback; and AMD’s own capacity trade-off, shifting shared production lines toward its supply-constrained CPUs. None of these is a demand-destruction story. All of them are timing stories.

Racks versus 8-GPU boxes: Helios at 40%, heading to 60%

The MI400 family splits cleanly by form factor. The top-end MI455X is built almost entirely for the Helios rack-scale system; the lower-end MI400 variants ship in conventional 8-GPU servers. For 2026, Helios racks account for roughly 40% of MI400-series shipments. In 2027, as 455X volume steps up sharply, that flips: Helios approaches 60%, with 8-card machines at about 40%. AMD is making the same transition Nvidia made — from selling cards and boxes to selling integrated racks.

Early Helios feedback is a genuinely mixed report card. On the plus side, the engineering fundamentals are in place: pooled memory, VRAM, and storage; virtualization; network tuning; full-rate switching within a single cabinet; monitoring and system-optimization software. The basics work. But early customers say overall performance and interconnect bandwidth aren’t yet pushed to the limit — the constraint sits with UALink and the switching layer, which matters most for very large model training and multi-GPU traffic across the network fabric. The expert’s assessment: if Helios can’t close that gap quickly, the largest models will run worse on Helios than on Nvidia’s B-series racks. He classed the known issues as normal for a just-shipped product, with Q3’s broader deliveries expected to surface more.

Two supply-chain notes. First, Wiwynn is effectively the sole manufacturer of the 455X — both Helios racks and 8-card machines. AMD opened the design package to Quanta, but Quanta’s Helios work hasn’t produced a shippable product; its focus remains conventional 8-GPU CSP servers. Wiwynn’s utilization is running near 80%, while the 8-card server lines across ODMs sit at 60–70% — a fair proxy for how measured AMD demand still is. Second, on Anthropic: despite public signals of interest, no actual order commitments are visible through any channel. Expect small validation purchases of MI450-class parts this year at most, with 2027 the earliest window for AMD to become a supplemental compute source there.

The takeaway

AMD’s 2026 is a supply story wrapped inside a timing problem. The units exist, the capacity is locked, and OCI is absorbing everything new AMD can build. What’s missing is the second anchor customer — Meta — and the proof point that Helios can hold its own against Nvidia’s racks at frontier-model scale. Both questions get answered in Q3. Watch the Helios evaluations, not the headline shipment number.

This update is based on an expert interview with a senior industry specialist . Figures are the expert’s estimates and should be treated as directional.

DeepSeek — the Chinese AI lab that rattled global markets and Silicon Valley’s assumptions about the cost of frontier intelligence — has completed the first external funding round in its history. The cap table is a snapshot of Chinese industrial power: Liang himself put in 20 billion yuan; Tencent added 10 billion; battery giant CATL contributed 5 billion; NetEase, JD.com and IDG Capital put in 3 billion each; and China’s national AI industry investment fund took a 1 billion yuan stake.

This is the same founder who once declared a principle of “no fundraising, no IPO, no commercialization.” So when DeepSeek recently held a nearly four-hour investor briefing, the industry paid attention — and a transcript of the session, spanning 118 questions and answers, has been circulating widely across Chinese media. The fullest version, published by Tencent Tech, is the most extensive public accounting yet of how Liang thinks about AGI, competition, chips, open source, and what kind of company DeepSeek actually wants to be.

Here are the most important things he said, and why they matter.

(Note: All quotes below are translated from the Chinese transcript; wording has been lightly edited for clarity while preserving meaning.)

1. Vision over money — and restraint as strategy

Liang opened with a point he returned to again and again: DeepSeek was not founded to make anyone rich.

“When we started this company, we never thought about how much money we’d eventually make, or going public, or anything like that. The first few dozen people — if they had thought that way, they wouldn’t have come,” he said. “We approached this with enormous goodwill toward the world. We believed it was useful for humanity. It was something beyond money.”

This sounds like standard founder mythology, except that Liang has operationalized it into what he explicitly calls a strategy: restraint.

“Restraint is a strategy. Sometimes you give something up to gain more elsewhere,” he explained. The logic runs like this: the prize of AGI is so large that fighting over slices of it now is a distraction. “AI is too big, the stakes are too large. As long as we succeed, the rewards will be enormous — take any small slice and it’s still enormous. So there’s no need to think about which part of the pie to grab or how to grab it.”

When DeepSeek’s user numbers exploded during last year’s Spring Festival (the global “DeepSeek moment”), the company didn’t rush to monetize, didn’t fight to lock in users, didn’t build a super-app. “We had no thought of becoming the next ByteDance or the next Tencent. None at all,” Liang said.

Most strikingly, he framed restraint as a probability-maximizing move rather than a moral one: “I have no doubt AGI will have enormous commercial value. But given that, my first priority isn’t how to take a bigger share. My first priority is how to increase the probability that we actually get there.”

Whether you read this as philosophy or as extremely long-term greed, it explains nearly every unusual decision DeepSeek has made — including open-sourcing its crown jewels.

2. The AGI roadmap: ladders, continuous learning, and the “singularity that isn’t”

Liang laid out the clearest version yet of DeepSeek’s technical worldview. AI progress, he said, is a staircase:

• Last year’s step was chain-of-thought (CoT) — reasoning as the unlock for higher intelligence.

• This year’s step is agents — expanding what models can do, built on top of CoT, which is built on top of language models. “Not a single step was wasted.”

• The next step is continuous learning — models that learn on the job over long horizons, the way a new employee becomes valuable after two months at a company. “AI can’t replace your employees today. But if it could learn continuously, like an employee who joins and learns for two months, then it could replace anyone. That’s what we’re missing.”

• After continuous learning comes the self-improvement “singularity” — a model that can develop its own next version. Liang immediately hedged: “It’s not really a singularity. It’s a gradual process, possibly a long one. We just habitually call it a singularity.”

• Finally, embodied intelligence — robots that do housework and care for the elderly.

The sequencing is the strategy: “If we solve continuous learning first, then the self-iterating singularity, then embodiment, the road gets easier — because you can use earlier technology to develop later technology.” In other words, get AI to accelerate AI research, and progress becomes nonlinear.

Equally telling is what DeepSeek won’t do. Video generation? “A good business, but it has nothing to do with the intelligence roadmap. We won’t do something just because it’s a good business.” World models? “Not the most important thing at this stage.” The company works only on what Liang calls the “main line” of AGI — and the next generation of models, he insisted, only deserves the name if it can learn continuously.

3. “Our only non-negotiable core interest”: the team

Ask a typical AI CEO about their biggest risk and you’ll hear about chips or regulators. Liang’s answer: keeping his researchers.

“Our greatest core interest is the stability of the team. It might be the only core interest. As long as I can keep the team stable, we will definitely build AGI. It’s that simple,” he said. “Money is not a problem. Resources are not a problem. Everything else is easily obtained. There’s only one thing we can’t compromise on.”

The new funding round, he admitted, substantially defused that risk — employees now hold meaningful options. But his theory of retention isn’t compensation-first: “As long as the most important, most senior people stay, others won’t leave even with smaller grants, because they didn’t come purely for money. Everyone wants to work in an environment where AGI can actually be built.”

His description of DeepSeek’s internal culture is unusual by the standards of both Chinese tech and frontier labs: no KPIs, no performance reviews, a “vision-driven” organization where the vision isn’t even written down. Employees are expected to keep roughly half their time unscheduled, for self-directed exploration. Overtime is discouraged — “research requires a relaxed environment; if you squeeze people tight, they can’t do research” — and possible anyway because the company is so focused there simply isn’t that much assigned work. Decision-making runs on consensus, not founder fiat, though Liang concedes that with headcount growing, more formal structure is now unavoidable.

4. Compute: the real bottleneck, and the “1/20th” narrative he wants to rewrite

Liang was blunt about the US-China gap: it is, at bottom, about compute.

“The biggest gap between us and the US is resources,” he said. “Almost all the differences you see — in talent, in model capability, in applications — can be attributed to differences in compute resources.” Even the talent gap, in his telling, is downstream of hardware: fewer GPUs means fewer experiments, which means less-trained researchers.

His now-famous framing of DeepSeek’s achievement: “We’re one to two years behind the US, and we did it with one-twentieth of the compute.” But he’s not satisfied with that story. “In the future we want to rewrite that narrative — use a fraction of their compute, but shrink the time gap to six months, three months. That’s the goal.”

On scaling, he sided squarely with the scalers — with a twist: “When Silicon Valley says scaling is hitting a wall, that’s true for Silicon Valley. For us in China, we’re nowhere near that point. We simply haven’t scaled to that degree.” DeepSeek’s model sizes, he revealed, are determined not by any belief that a given size is “enough,” but by a cold calculation of what the available resources can support. With the new war chest, the strategy is simple: buy as many GPUs as possible at reasonable prices — so hard, he joked, that “if we manage to spend 20 billion yuan this year, our procurement team has done a superhuman job.”

5. Chips: “NVIDIA is digging its own grave”

Some of the most headline-worthy remarks concerned the chip war — and they should be read as those of an interested party, but a well-informed one.

Liang argued that CUDA’s moat is “being rapidly dismantled,” partly because AI itself now makes it far easier to build alternative software ecosystems. He disclosed that DeepSeek trained V3 on NVIDIA hardware without relying on NVIDIA’s software ecosystem — instead using its own high-level compiler, TileLang, for nearly everything above the silicon.

He was effusive about Huawei: its upcoming 950 supernode, he claimed, can fully substitute for NVIDIA’s GB200/GB300 on both performance and price. “Four Huawei cards match one NVIDIA card,” he said, summarizing the hardware gap as “4x plus two years” — but with the ecosystem gap effectively closing. Within a year, he predicted, the industry will have proof that the domestic-chip ecosystem “has no problems at all.” The only real constraint is manufacturing capacity, which he expects to remain binding for two to three years but not five.

His most quotable line: “On this point, NVIDIA is digging its own grave. In a normal commercial environment where we could buy NVIDIA cards, domestic substitution would be hard. But when NVIDIA’s cards can’t be bought, everyone is forced onto domestic chips.”

6. Competition: “Whoever wants to take more will be beaten by whoever takes less”

Liang’s view of the global model race is unsentimental. Anthropic’s edge over OpenAI right now? “Phase-specific, not long-term. OpenAI and Google will most likely take turns pulling ahead.” China’s many foundation-model companies? “Too many. The US has maybe three. This will definitely consolidate” — perhaps to two large and two small players globally, competing only on cost and timing. “No one will earn windfall profits. Those who control costs well earn a bit more; those who don’t, earn a bit less. That’s all.”

China’s structural role, in his telling, mirrors its position in manufacturing: largest capacity, cheapest power, and ultimately “Chinese-made AI will be systematically cheaper, the way Chinese services are cheaper in other industries.”

And then the philosophical thrust that ties back to his opening theme: “Those who take more will be defeated by those who take less. You don’t even have to actually take more — if your vision is to take more, you’ve already lost.” OpenAI, he said, “thought from the beginning it could monopolize the world. In reality it will face many challengers — including from China, because Chinese players are willing to take less while providing the same service.”

7. Open source: “I can’t see any benefit to closed source”

On open source, Liang was categorical — and this is perhaps the answer global developers most wanted to hear.

“We will open source, and our strongest models will likely be open sourced too. I can’t see what benefit closed source brings,” he said, taking an explicit swipe at ByteDance’s closed approach. “Even when you open source everything, the barrier to actually using it well is extremely high. Using it is hard; using it at low cost is very, very hard.”

Open sourcing, he argued, has zero negative impact on DeepSeek’s revenue, and the open weights are identical to what DeepSeek itself deploys — no crippled community versions. He welcomes others deploying DeepSeek models to compete with DeepSeek itself, and says the company actively helps rivals — naming Alibaba, Zhipu and Moonshot AI — do better. “We lose nothing. We were open source anyway.”

And his counterintuitive punchline: “We haven’t received less because we open sourced. If anything, we’ve gained. It looks counterintuitive, but that’s how it is.”

8. The business: reasonable profit, not maximum profit

For all the anti-commercial rhetoric, Liang was clear that DeepSeek is a company, not a monastery: “We definitely intend to commercialize. We have to survive — the government won’t give me a cent.” The company already has consumer users and B2B revenue, and he sketched a path where “a few hundred million dollars” of B2B revenue this year, growing next year, puts net profit within reach. “Worst case, if technology froze today and all we did was sell APIs well, that alone could support a listed company.”

But pricing follows the restraint doctrine. The API, he said, is priced for a “reasonable profit” — roughly a ten-month payback on purchased hardware — even though he openly acknowledged demand is price-inelastic: “If we wanted to maximize profit, we’d price higher. Double the price and token consumption barely changes.” He recounted overriding his own team’s initial high pricing on one model and cutting it to a quarter of the original level — “and everyone was happy.”

On product priorities: general-purpose agents first, coding agents above all; vertical agents (finance, medicine) later. Multimodality will arrive natively in V4 and beyond, but “it’s a component, not the main line of intelligence itself.”

One more operational detail that deserves attention: half of DeepSeek’s core researchers currently spend their time on data labeling. “At this stage, solving AI comes down to labeling data,” Liang said, noting that high-end data annotation in China carries no cost advantage over the US — a genuine constraint on the post-training pipeline, where he also believes hallucination can eventually be engineered away.

Why this briefing matters

Strip away the specifics and what remains is a coherent, unusual theory of how to win the most consequential technology race in history: constrain your ambition to a single technical main line, give away your best work, refuse to maximize prices, spend everything on compute and talent stability, and trust that the pie is big enough that being the team that gets there matters more than the share you grab along the way.

There are reasons for skepticism. The “no commercialization” principle has now been revised once; the new investors — Tencent, CATL, JD.com — will eventually want returns; a valuation north of $50 billion has a way of rewriting cultures that pride themselves on having no KPIs. And some of the chip claims (four-to-one parity, full GB300 substitution) await independent verification.

But Liang ended on a note that explains why so many talented researchers — and now, so much capital — are willing to bet on his version of restraint: “Many companies became great because they pursued something beyond profit. That pursuit didn’t hurt their commercialization. It made them commercialize better.”

For a global audience trying to understand what DeepSeek is, this four-hour session offers the best answer yet: not a cheaper OpenAI, not a state project in disguise, but a lab betting that in the age of AGI, the most aggressive strategy of all is to want less.

Source: This article is based on a transcript of Liang Wenfeng’s remarks at a recent DeepSeek investor briefing, compiled and published by Tencent Tech (腾讯科技). Quotes have been translated from the original Chinese and lightly edited for readability. Figures on the funding round are as reported in that transcript.

MediaTek never named the customer. But the supply chain has: TrendForce, J. P. Morgan, and Morgan Stanley all point to Google’s TPU. Treat that attribution with the appropriate asterisk — it is a near-unanimous market judgment, not a confirmed fact. What is confirmed is that a company known mostly for smartphone chips just booked a hyperscaler ASIC order large enough to equal roughly 42% of its Q4 2025 revenue in a single quarter.

The interesting question isn’t whether MediaTek replaces Broadcom. It doesn’t, and the numbers below show why. The interesting question is why Google decided its most strategic chip program needed a second designer at all — and what gets rearranged when a hyperscaler starts unbundling a turnkey relationship.

The TPU roadmap now has two names on it

Some context before the analysis. Based on Morgan Stanley’s supply chain checks, here is how the current and upcoming TPU generations map to design partners:

Every generation before this — from the original 2015 inference chip through v6 Trillium — ran through Broadcom. The codenames and partner assignments above come from sell-side checks, not from Google, and the caveat matters: research desks don’t even agree on internal naming. But the packaging bookings behind them are harder to fake. Morgan Stanley sees MediaTek’s TSMC CoWoS-S reservations jumping from 40k wafers in 2026 to 180k in 2027, while Broadcom’s Google-related bookings sit at 365k. Nobody books 180,000 advanced-packaging wafers for a pilot project.

The chart makes the story legible at a glance: in 2026, MediaTek’s ZebraFish is a modest sideline next to Broadcom’s Ironwood and SunFish volumes. By 2027, the two partners are shipping at near parity in unit terms — roughly 4.2 million implied units for Broadcom against 4.0 million for MediaTek — even though the value per chip still tilts heavily toward Broadcom.

Why a second designer, really

The lazy answer is “cost,” and it’s at best half right. Start with what the partnership actually is. Google has always defined the TPU’s compute architecture itself; the ASIC partner’s job is turning that architecture into manufacturable silicon and a deployable system. But within that frame, the commercial models differ enormously. At one end sits turnkey, where the partner handles physical design, IP, tape-out, packaging, testing, and procurement — and charges for all of it. At the other sits COT, where the customer owns most of the design and the partner is closer to an execution service. J. P. Morgan characterizes the Google–MediaTek arrangement as “semi-COT”: Google keeps the compute die, MediaTek handles the I/O die, SerDes, back-end physical design, and packaging integration.

That framing explains motive number one. Broadcom’s model with Google sits near the turnkey end, and turnkey pricing has a scaling problem — for Google. When TPU volumes were in the hundreds of thousands, a service markup on wafers, HBM, and substrates was a rounding error. J. P. Morgan now models total TPU shipments reaching 8 million units in 2027. At millions of units, even a modest percentage markup on pass-through component costs compounds into hundreds of millions of dollars a year. Google has every incentive to claw back procurement and product definition, and a second partner operating under a thinner semi-COT model is the lever.

But — and the Morgan Stanley report is refreshingly honest here — the savings aren’t guaranteed. Google wants to source bulk components like HBM directly, yet spot HBM prices now run above the contract prices Broadcom locked in years ago. Unbundling the turnkey stack may trim service fees while surrendering Broadcom’s accumulated purchasing scale. In the near term, the second source might not be cheaper at all.

Which points to the stronger motive: delivery risk. A frontier AI chip is hostage to TSMC leading-edge capacity, HBM supply, CoWoS packaging, ABF substrates, SerDes validation, power, and cooling — miss any one and a data center slips its schedule. Broadcom disclosed over $30 billion in AI semiconductor orders in a single quarter of fiscal 2026, with visibility stretching to 2028. When your sole design partner is that oversubscribed, concentration itself becomes the risk. A second qualified path from architecture to packaged silicon is insurance, and at Google’s spend level, insurance worth paying for even if it saves nothing.

The third motive is product segmentation. Training silicon lives and dies on HBM bandwidth, interconnect, and packaging complexity; inference at scale is a cost-per-token game. Deutsche Bank expects the split to follow exactly that line: MediaTek takes the efficiency-oriented, inference-leaning SKUs while Broadcom keeps the complex, higher-value parts. The Morgan Stanley data agrees — SunFish carries a larger die and higher chip value than ZebraFish, and the inference-focused TriggerFish variant may even support TPU leasing. MediaTek can win on unit count and still trail badly on revenue share.

What this actually changes

For MediaTek, nearly everything. Twenty billion New Taiwan dollars of quarterly mobile-chip revenue doesn’t rerate a stock; $2 billion a quarter of hyperscaler ASIC revenue does — J. P. Morgan models roughly 2.5 million chips and $10.8 billion of ASIC revenue in 2027, with implied ASPs jumping from about $4,300 to over $12,000 in the following generation. Gross margin will fall, perhaps to 42% by 2028, and the reflex is to read that as deterioration. Resist it. The design costs are already sunk and the incremental opex is thin, so J. P. Morgan sees operating margin climbing from 17% to as much as 25%, with data center ASICs contributing perhaps 70% of operating profit by 2028. The real test is whether MediaTek’s SerDes, die-to-die, and 3.5D packaging capabilities — built quietly since 2018 — can land a hyperscaler that isn’t Google. One concentrated customer is a windfall; two is a platform.

For Broadcom, the loss is precise: exclusivity and the turnkey premium, not the customer. Its 2027 Google-related packaging bookings still exceed MediaTek’s, and every TPU rack — including ones with MediaTek silicon inside — still needs Broadcom’s switch chips and networking, which its Google supply agreement explicitly covers. The right metric is no longer TPU unit share but gross profit captured per gigawatt of TPU capacity built. Unit share and profit share are different games, and Broadcom still leads the one that pays better.

For everyone else, the signal is the precedent. The largest custom-silicon buyer on earth has demonstrated that the turnkey ASIC model is negotiable — that design, procurement, packaging, and networking can be pulled apart and priced separately. Amazon, Meta, and Microsoft are all watching. The moat in AI ASICs was never one chip. It was the system around it, and Google just proved even that can be split.

This shift was crystallized in Huawei’s launch of Tao’s Law—a bold articulation that advanced packaging innovation could help circumvent front-end limitations by extracting maximum performance from available chip technologies through sophisticated integration techniques.

The message is clear: if China cannot access the most cutting-edge chip fabrication processes, it will compensate by becoming world-class at assembling multiple chips into high-performance systems through advanced packaging. This strategic reorientation has triggered a wave of investment, capacity expansion, and technical development across China’s packaging industry, particularly in 2.5D and 3D technologies like CoWoS (Chip-on-Wafer-on-Substrate)—the critical enabler for high-performance AI processors.

To understand where this race actually stands, I drew on a recent expert interview conducted by a Chinese research firm that provides rare visibility into the competitive dynamics, technical challenges, and market realities shaping China’s advanced packaging sector. The insights reveal a market in rapid transformation, with established players racing to expand capacity while newcomers struggle to break through formidable barriers. The stakes are enormous: control over advanced packaging could determine which companies—and which nations—lead the AI revolution.

The Two-Track Development Model

Since 2024, China’s advanced packaging landscape has crystallized around two distinct groups of players, each pursuing different strategies but converging on the same goal: mastering 2.5D and 3D packaging technologies.

The first tier consists of established, publicly-traded packaging giants—SJ Semiconductor, Tongfu Microelectronics, JCET (Jiangsu Changjiang Electronics Technology). These incumbents are methodically upgrading their capabilities with a clear hierarchy of priorities. CoWoS-like 2.5D production lines come first, followed by Fan-Out technology upgrades for consumer and industrial applications, and finally, forward-looking investments in panel-level packaging.

The second tier comprises smaller or newly-established advanced packaging firms. Well-capitalized players like 江苏华天 and 甬矽are building full-process 2.5D/3D production lines. Meanwhile, companies with more limited resources are carving out niches in specific segments of the CoWoS process—some focusing exclusively on the OS (On-Substrate) portion of the workflow.

From 2023 onward, the overwhelming majority of capital and human resources has flowed into 2.5D/3D packaging, with Fan-Out and panel-level packaging receiving secondary attention. This laser focus reflects market realities: AI chips demand advanced packaging, and China’s domestic AI chip designers need local packaging capacity.

Why There’s No Shortcut to CoWoS-L

For companies eyeing the advanced packaging market, a common question arises: can they leapfrog directly to CoWoS-L, the large-format variant designed for the most powerful AI accelerators, and achieve a “bend-overtake-on-the-curve” advantage?

Industry experts are unequivocal: no such shortcut exists. CoWoS is fundamentally a composite of multiple packaging technologies, requiring mastery of four core competencies: wafer-level packaging for bumping processes, flip-chip technology for die attachment, middle-of-line (MEOL) wafer fabrication processes for silicon interposer manufacturing, and redistribution layer (RDL) technology. CoWoS-L adds even more demanding requirements for RDL or Fan-Out routing, plus MEOL processing for silicon bridges.

Every company must first master CoWoS-S before progressing to CoWoS-L. Across mainland China, all manufacturers are following the “S-first, L-later” pathway without exception.

This sequential approach makes strategic sense. While CoWoS-L represents the future for ultra-high-performance computing chips, CoWoS-S won’t disappear. Smaller, lower-power chips for inference and edge computing applications will continue to require CoWoS-S packaging, much like how traditional packaging technologies persist despite the emergence of advanced alternatives.

The Triple Barrier to Entry

For new entrants contemplating the advanced packaging market, the barriers are formidable and multifaceted.

Capital requirements form the first hurdle, though surprisingly, they account for only about 20% of the challenge. Building a CoWoS-S production line capable of producing approximately one million units annually—using chips comparable to Huawei’s 910-series as reference—requires at least 1.2 billion RMB in investment for facilities and equipment combined.

Time costs represent the most daunting barrier, weighing in at roughly 60% of the total challenge. The journey from factory construction to customer qualification and mass production is brutally long. Industry benchmarks tell the story: SJ Semiconductor took three years from 2020 to late 2023 to reach volume production. Tongfu Microelectronics needed three years from pilot line to mass production (2021-2024). JCET required approximately four years (2021-2025). The average timeline to achieve meaningful production volume is at least three years—an eternity in the fast-moving semiconductor industry.

Customer access completes the triumvirate of barriers, accounting for the remaining 20%. The mainland China market for 2.5D packaging currently centers almost exclusively on AI GPU chips. For the foreseeable five-year horizon, mainland packaging facilities will serve only domestic chip design companies—AMD, NVIDIA, and other international players won’t be routing production through Chinese facilities.

The pool of potential customers is remarkably small. Huawei can generate nearly one million units of annual packaging demand. Cambricon, Hygon, and Baidu’s Kunlun each require several hundred thousand units. Smaller players like Tianshu Zhixin, Biren, and Suiyuan operate at the tens-of-thousands level. These customers overwhelmingly prefer working with top-tier OSATs (Outsourced Semiconductor Assembly and Test providers), making customer acquisition extraordinarily difficult for newcomers.

The Current Competitive Landscape

In the CoWoS-S arena, the mainland China market has stratified into three clear tiers.

The first tier consists of companies achieving genuine CoWoS-S mass production with meaningful 2025 revenue. Ranked by shipment volume, revenue, and process capability: SJ Semiconductor leads with approximately one million units of annual capacity in 2025—the highest in China. Tongfu follows with 500,000 units of annual capacity by end-2025. JCET rounds out the tier with 300,000 units by end-2025.

In terms of specifications, SJ Semiconductor’s mainstream products feature 1+6 configurations (one compute die plus six HBM stacks) with 2+8 configurations already in development. Tongfu’s mainstream is 1+4 with some 1+6 capability. JCET focuses primarily on 1+4 with partial 1+6 capacity.

The second tier includes companies with production lines installed but not yet in mass production, currently in customer qualification phases. Jiangsu Huatian and Yongsi Electronics occupy this space. Jiangsu Huatian’s Nanjing facility has 300,000 units of annual capacity in 2025, while others remain at the tens-of-thousands level.

The third tier encompasses companies in the planning, fundraising, or construction phases, including Biwin ‘s subsidiary Xincheng Hanqi and Quliang.

CoWoS-L development lags behind CoWoS-S overall, reflecting lower demand from domestic design companies for large-format products. In the first tier for L-type lines: SJ Semiconductor has 300,000 units of capacity in 2025, JCET approximately 300,000 units, and Tongfu maintains a pilot line with under 100,000 units. The second tier includes Huatian’s pilot line (under 100,000 units) and Yongsi’s planned 100,000 units, with others yet to establish L-type capabilities.

The Customer Diversification Imperative

An interesting dynamic is emerging around customer concentration. While SJ Semiconductor has enjoyed first-mover advantage and exclusive supplier status with Huawei, the latter is now actively cultivating alternative packaging capacity—including at Quliang—to reduce dependency on a single supplier.

Currently, Quliang primarily produces consumer chips for mobile Wi-Fi, video decoding, and SoC applications. Starting in the second half of 2025, the major customer began gradually shifting some large-format chip packaging to Quliang, initially focusing on FPGA and CoWoS back-end processes. Quliang lacks mass production capability for CoWoS front-end processes—no equipment installation or production line integration has occurred—keeping it firmly in the third tier. However, by the second half of 2027, Quliang could potentially reach second-tier status.

Huawei’s diversification trend is forcing market leaders to expand their customer bases. SJ Semiconductor, facing potential order redistribution, must court new clients.

Fortunately, the company’s technical position makes this transition manageable. With industry-leading yields and extensive experience in high-end chiplet stacking and large-format packaging, SJ Semiconductor has attracted proactive inquiries from multiple chip design companies in the Yangtze River Delta region. The company has simply prioritized maintaining its primary customer relationship rather than aggressively pursuing new business—but when the time comes, external customer acquisition faces minimal resistance.

Meanwhile, Tongfu and JCET are actively qualifying additional customers including Suiyuan, Moore Threads, and Biren. The emerging pattern suggests that mature, older products will gradually transition to captive or affiliated packaging facilities, while cutting-edge new products will continue prioritizing SJ Semiconductor initially, eventually evolving into dual-source supply arrangements.

Supply-Demand Dynamics Through 2027

Despite aggressive capacity expansion, the advanced packaging market is not heading toward oversupply or price wars—at least not through 2027.

The industry currently operates in supply-constrained mode. Domestic AI chip design companies hold wafer inventories sufficient to support approximately four million high-end chip units annually, and 2026 demand will very likely exceed four million units. Even if all first- and second-tier manufacturers reach full production in 2027, they will struggle to satisfy four million units of high-end chip demand. Capacity will remain tight.

Outside of the bumping segment—which faces overcapacity and intense competition—flip-chip and 2.5D/3D packaging show no signs of commoditization. Barring material-driven price increases, the industry won’t initiate voluntary price reductions in 2026. A full-year price war won’t materialize in 2027, though 2028 remains uncertain.

The Equipment Bottleneck

For packaging facilities attempting to scale with major customer support, the critical constraint isn’t technology or customer commitment—it’s imported equipment delivery times.

Molding equipment requires approximately 12 months for delivery. DISCO grinding and dicing equipment needs about 8 months. Chip-on-wafer die bonding equipment demands a staggering 18 months. Even with major customer intervention to expedite orders, these lead times cannot be compressed.

This equipment bottleneck defines realistic expansion timelines. A facility beginning construction in the second half of 2025 could, at best, complete a pilot line by mid-2027 with annual capacity around 100,000 units. Subsequent expansion would take roughly another year, following a “produce-while-expanding” model.

Customer Relationships Trump Technology—For Now

In the current market phase, securing major customer relationships matters far more than technical prowess. Industry insiders estimate that major customer binding accounts for 80% of competitive advantage, while proprietary technology represents only 20%.

Companies locked into relationships with large customers can more easily secure substantial investments and continuously iterate on high-end processes. The industry has already formed fixed customer-plus-packaging partnerships, creating a calcifying competitive structure.

However, this imbalance will gradually correct. For standard 1+4 configurations, yield rates across manufacturers have already converged. For more advanced 1+6 and higher configurations, yield improvements follow customer chip design iteration cycles. Packaging facility yields improve in lockstep with customer chip design evolution—it’s merely a matter of timing. Over the next one to two years, yield rates among mainland China’s leading manufacturers should progressively equalize.

As technical capabilities converge, competitive differentiation will shift toward operational efficiency, customer service, and the ability to support next-generation packaging architectures. The companies that can balance customer relationships with genuine technical innovation will emerge as long-term winners.

Conclusion: A Market Taking Shape

China’s advanced packaging industry is in a critical formation phase. The technology barriers are real, the capital requirements substantial, and the time horizons long. Yet the market is developing with remarkable speed, driven by insatiable demand for AI computing power and the imperative of supply chain localization.

The competitive landscape through 2027 appears relatively stable: established leaders will maintain advantages, capacity will remain constrained, and pricing will hold firm. But beneath this surface stability, important shifts are occurring—customer diversification, yield convergence, and the gradual emergence of second-tier players.

For observers of the global semiconductor industry, China’s advanced packaging sector offers a fascinating case study in how markets develop under conditions of technological challenge, geopolitical pressure, and explosive demand. The next two years will determine whether China can build genuinely world-class advanced packaging capabilities—or whether the barriers to entry will prove too formidable for all but a handful of winners.

The race is far from over, but the starting positions are now clear.

To understand why this matters, we need to grasp what makes indium—and specifically indium phosphide (InP)—so critical to modern telecommunications.

The Invisible Material Powering the Data Economy

Indium phosphide substrates are thin wafers that serve as the base material for growing the semiconductor layers used in optical chips—the devices that convert electrical signals into light for transmission through fiber optic cables, and back into electrical signals at the receiving end. Think of the substrate as the foundation upon which the active optical components are built.

Why InP specifically? The answer lies in physics. Optical fiber has two “windows” where signal loss is minimal: around 1,300 nanometers and 1,550 nanometers wavelength. InP-based materials naturally emit and detect light at precisely these wavelengths, making them ideal for long-distance, high-speed data transmission. Silicon, by contrast, cannot efficiently emit or detect light at these critical wavelengths—it’s simply the wrong material for the job.

As data centers scale to handle AI workloads and cloud computing demands, the need for faster optical interconnects has exploded. Today’s 800G and 1.6T optical modules—which enable data transmission at speeds that would cause silicon-based electronics to overheat and fail—all depend on InP substrates as their foundation material.

The market numbers tell the story: the global InP substrate market was valued at approximately $150-200 million in 2024, projected to reach $230-240 million in 2025, and $540 million by 2026. While these figures seem modest compared to the broader semiconductor industry, they represent a critical bottleneck. The optical module market built on these substrates is 50-100 times larger, and the downstream data center infrastructure market is another 50-100 times beyond that.

China’s Dominant Position

Here’s where geopolitics enters the picture: China controls approximately 70% of global primary indium production, producing around 750 tons annually from its 1,000-ton global total. Add recycled indium to the equation, and China’s total annual production reaches roughly 1,050 tons—over 80% of the world’s supply.

This dominance isn’t accidental. Indium doesn’t exist as an independent ore; it’s a byproduct of processing zinc, lead, and tin. China’s massive base metals industry naturally yields substantial indium output, creating a structural advantage that would take competitors years and billions of dollars to replicate.

The export controls implemented in February 2025 don’t ban indium exports entirely, but they subject them to strict licensing requirements. More significantly, China has signaled enforcement of “origin principles”—meaning any final product using Chinese-origin indium or gallium, regardless of how many times it changes hands, cannot be used for military applications. This effectively extends Chinese regulatory reach deep into global supply chains.

The Supply Chain Squeeze

The immediate impact has been felt most acutely by the established players in InP substrate manufacturing. The global market is dominated by a handful of companies: Japan’s Sumitomo Electric (approximately 50% market share in 2024), Japan’s JX Nippon Mining (5%), Germany’s Freiberger, and the American company Coherent (which acquired II-VI). Chinese domestic producers like Yunnan Xinyao (a subsidiary of Yunnan Germanium) and Guangdong Xian Dao have been rapidly scaling production.

For the Japanese and Western manufacturers, the math is uncomfortable but not immediately catastrophic. Global non-Chinese indium production stands at approximately 350 tons annually. Producing 1-1.2 million pieces of 2-inch equivalent InP substrates globally requires less than 5 tons of pure indium. Even accounting for production waste and material loss, total demand might reach 10 tons. The optical module industry’s total indium requirement sits somewhere between 50-100 tons annually—manageable within available non-Chinese supply, especially considering indium’s other applications like transparent electrodes.

However, this analysis assumes perfect market efficiency and ignores several complicating factors. First, establishing reliable non-Chinese indium supply chains takes time, requiring new supplier relationships, quality verification, and production adjustments. Second, prices have already responded to the tighter supply situation. Third-inch InP substrate prices, which stood at $310-350 per piece in early 2025, have risen approximately 20% to around 3,000 RMB (roughly $450) per piece by late 2025.

The Technology Race Intensifies

Perhaps more significant than the immediate supply disruption is how export controls are accelerating technological and geographic shifts in the industry.

Japanese manufacturers, particularly Sumitomo, have responded by dramatically increasing production capacity—expanding by 40-100% according to industry sources. This expansion has helped stabilize international market prices, suggesting that Sumitomo, JX, Coherent, and Freiberger have successfully absorbed much of the supply gap through increased output and efficiency improvements.

Meanwhile, Chinese domestic producers are racing to capture market share. Companies like Yunnan Xinyao have been aggressively scaling production since 2021, with furnace counts approaching 100 units by 2024. While Chinese manufacturers still lag behind Japanese competitors in yield rates—industry sources suggest Chinese producers achieve perhaps one-quarter to one-fifth the customer-end yield of Sumitomo products—the gap is narrowing through continuous production experience and process refinement.

The yield differential is crucial to understanding competitive dynamics. According to an expert interview transcript obtained by Compute’s Research, achieving the same output might require 10 pieces of one supplier’s substrate versus just 2 pieces from a superior supplier. This isn’t merely about crystal growth; it encompasses the entire production environment, equipment quality, and process control. For InP substrates, where specifications are more demanding than for gallium arsenide, these factors create enormous performance variations between suppliers.

Strategic Maneuvering and Workarounds

The export control regime has also sparked creative strategic responses. Coherent has reportedly established operations in China, potentially planning to use domestically-produced 6-inch substrates for preliminary processing before exporting semi-finished products to overseas facilities for final manufacturing. If successful, this approach could circumvent export restrictions while leveraging China’s cost advantages and material access.

This highlights a broader pattern: rather than completely severing supply chains, companies are developing hybrid models that navigate the regulatory landscape while maintaining operational efficiency. The result is a more complex, fragmented global production network—less efficient than the previous integrated model, but more resilient to single-point regulatory interventions.

The Broader Context: Technology Decoupling

China’s indium export controls didn’t emerge in isolation. They represent one move in an ongoing technology competition between China and Western nations, particularly the United States. After the U. S. restricted ASML’s advanced lithography equipment exports to China, Chinese officials signaled stronger controls on critical materials exports. The indium restrictions followed earlier controls on gallium and germanium.

The timing is also significant. As artificial intelligence drives explosive growth in data center construction and high-speed optical interconnects become essential infrastructure, controlling access to InP substrates provides leverage over a chokepoint in the global technology stack.

For Chinese policymakers, the calculus is straightforward: leverage dominant positions in raw materials and mid-stream processing to secure strategic advantages while domestic companies climb the value chain. For Western policymakers, the indium situation exemplifies the vulnerabilities created by concentrated supply chains and the challenges of securing access to materials that seemed mundane until they became critical.