Shanghai Iluvatar CoreX: China's GPU Gamble Goes Public

I. Introduction & Episode Roadmap

On the morning of January 8, 2026, a company with no founder, no controlling shareholder, and no profit from its actual product rang the bell at the Hong Kong Stock Exchange. The banners said 上海天数智芯半导体股份有限公司 Shanghai Iluvatar CoreX Semiconductor Co., Ltd. The people standing behind the gong were, for the most part, not the people who had started it. The man who founded the company had walked out in a dispute with his own investors. The chairman who replaced him had been taken away for investigation and lost contact with the business. The executive holding the gavel was an accountant by training, a private-equity investor by career, and a chip designer by neither.

The market did not care. Shares opened up 31.5%, and within minutes the company was worth more than HK$47 billion — roughly US$6 billion — for a business that had, over a decade of existence, never earned a single yuan of profit from selling chips.¹²

That gap — between what the company has proven and what the market has decided it is worth — is the whole story. It is also a reasonable proxy for the entire Chinese AI-semiconductor trade of the mid-2020s.

Here is what Iluvatar CoreX genuinely is, stripped of promotional language. It designs general-purpose graphics processing units — GPGPUs, the class of chip that does the enormous parallel arithmetic behind modern artificial intelligence. It was the first company in China to mass-produce a domestically designed GPGPU for training (teaching an AI model from raw data), and the first to mass-produce one for inference (running the finished model to answer a user's question), and it did both on a 7-nanometre process node.³² Those are real engineering firsts in a country that, until very recently, had no credible domestic answer to Nvidia at all.

Here is what it also is. A company that lost money in every year of its disclosed track record. A company whose much-celebrated "return to profitability" in the first half of 2026 turned out, on inspection of its own interim disclosure, to rest almost entirely on an unrealized paper gain from someone else's stock-market listing — an equity stake it is contractually forbidden from selling for a year.⁴ Strip that accounting gain out and the chip business lost more money in six months of 2026 than it had lost in any full year before.

Both of those things are true at once, and neither cancels the other. That is what makes this a story worth two hours rather than two paragraphs.

The Road Ahead

The narrative runs in roughly the order the company lived it. It begins in Nanjing in December 2015 with an ex-Oracle engineer and a strategy argument he lost — an argument that set the company's product direction and cost it its founder. It moves through the leadership vacuum that followed, including a chairman's disappearance into one of the largest corruption investigations in the modern history of Chinese semiconductors.

From there it turns technical, but not painfully so: what Iluvatar actually built, why the decision to be compatible with Nvidia's software rather than to fight it was the single most consequential product call the company ever made, and why the fact that its wafers come from 中芯国际 SMIC rather than TSMC sets a ceiling on everything else.

Then the financials, which tell a strange dual story of revenue compounding at extraordinary rates while losses compound alongside them — a structural feature of this entire industry cohort, not a peculiar failure of one management team.

Then 2026 itself: the listing, a HK$7.07 billion follow-on placement six months later, a bonus share issue, and a reported order from 字节跳动 ByteDance that, if it converts, would be the single most important commercial validation any of China's independent GPU startups has ever received. And, sitting immediately beside that good news, the profit figure that does not mean what the headlines said it meant.

Then the war game: the "four little dragons" of Chinese domestic GPUs all racing to the public markets at once, 华为 Huawei's Ascend line looming over all of them, 寒武纪 Cambricon setting a profitability bar none of them have cleared, and Nvidia — restricted, diminished, but still holding an estimated 70% of China's AI accelerator market — remaining the yardstick every customer measures against.

Then governance, ownership, and incentives, where the absence of a controlling shareholder is not a technicality but a defining fact. Then the bull and bear cases, run through Porter's Five Forces and Hamilton Helmer's 7 Powers. And finally, the two or three numbers that will actually settle the argument over the next several quarters — because the last two years of stock performance will not.

Start where the company started: with a founder who did not stay.

↑ Back to Top

II. Founding, the Founder's Exit, and Why That Still Matters

In December 2015, in Nanjing, a man named 李云鹏 Li Yunpeng registered a semiconductor company.⁵ He had been an R&D director at Oracle — a systems and database background, not a graphics one — and he was betting his career on a proposition that in 2015 sounded either visionary or delusional depending on who you asked: that China would eventually need its own accelerated-computing silicon, and that a startup, not a state institute, could build it.

Consider what 2015 looked like from inside China's chip industry. Nvidia's CUDA software platform had already been in the field for eight years and had quietly colonized every serious machine-learning lab on earth. The deep-learning boom was underway but not yet a national-security matter. There were no US export controls on AI accelerators. Any Chinese lab that wanted the best GPU in the world simply bought one. Founding a domestic GPU company in that environment was not an act of import substitution — the imports were freely available and excellent. It was an act of faith in a future nobody had announced yet.

That faith was, in hindsight, correctly placed. Li's timing on the market was superb. His timing on the company was not.

The Argument That Cost a Founder

Less than a year after the company's first chip taped out — that is, after the design was finalized and sent to a foundry to be physically manufactured, the moment a chip startup stops being a PowerPoint and becomes an engineering organization — Li Yunpeng and the company's investors fell into a strategic disagreement they could not resolve.⁵⁶

The substance of the dispute is worth sitting with, because it explains the company that exists today. Li wanted a broader path: a wider product aperture, more optionality about what the silicon would be used for. The investors wanted focus, and specifically they wanted a dedicated general-purpose GPU — the head-on, Nvidia-shaped product. Not a domain-specific accelerator tuned to one narrow workload, not a family of adjacent chips, but the hard, expensive, capital-devouring thing itself.

The investors won. Li left in 2021.⁵

Founders lose these arguments all the time, and the outcome is not automatically bad. In this case the investors' strategic instinct arguably proved right: the general-purpose GPU is precisely the product category that US export controls would make scarce and strategically precious within eighteen months of Li's departure. A narrower or more diffuse product line would have left the company positioned for a market that never materialized.

But the manner of the resolution matters as much as its content. This was not a founder retiring, or moving to a chairman's role, or handing over to a lieutenant he had trained. This was an investor bloc overruling the person who started the company on the most fundamental question a chip company faces — what to build — and the founder leaving as a consequence. That establishes something about where authority in this organization actually sits, and it has never been reversed.

The Unigroup Chairman and the Investigation

The man who took over as chairman in May 2021 carried a resume that, at the time, looked like an enormous coup.⁵

刁石京 Diao Shijing had been co-president of 紫光集团 Tsinghua Unigroup — for years the most aggressive, most heavily financed, most politically connected vehicle in Chinese semiconductors, the group that had at various points attempted to buy or build its way into memory, packaging, and mobile chipsets on a scale nothing else in the country matched. Recruiting Diao was a signal to the market that Iluvatar had graduated from startup to national-project adjacency.

Under his chairmanship the company shipped the thing it had been arguing about: 天垓 Tiangai Generation 1, China's first mass-produced domestically designed training GPU, in 2021.³² Whatever else is true about the Diao period, that is a genuine execution milestone, and it happened on his watch.

Fifteen months later, in July 2022, Diao Shijing was taken away for investigation.⁵⁶ He was swept up alongside 赵伟国 Zhao Weiguo, Unigroup's founder, in the graft investigation that consumed the group's leadership. He lost contact with Iluvatar. He did not come back.

Think about what that means operationally for a company at that stage of life. Iluvatar in mid-2022 was a pre-revenue-scale business — it would report only RMB 189 million of revenue for the full year — attempting the most technically demanding product category in commercial electronics, in a country where the government's attitude toward its own chip-investment establishment had just turned punitive.⁷ And its chairman simply stopped existing, from the company's point of view, without warning.

The Accountant Who Ended Up Running a GPU Company

Into that vacuum stepped 盖鲁江 Gai Lujiang, who became chairman and CEO in 2023.⁵

Gai was not a chip person. He was an audit and private-equity person: time at PwC, time at Deloitte, then an investment platform linked to the 中植 Zhongzhi group — itself a name that would later become synonymous with distress in China's shadow-banking sector. He arrived at the top of Iluvatar not as a technologist promoted for technical judgment but as an investor-side board representative filling a hole.

This is not an argument that a non-engineer cannot run a semiconductor company. Plenty have, and the discipline an accountant brings to a category famous for burning capital may in fact be part of why Iluvatar's losses have stayed far below some peers'. But it is an argument against a particular kind of narrative shortcut. There is no visionary founder-technologist at the helm of this company. There is no Ren Zhengfei figure, no Jensen Huang analogue, no person whose personal technical conviction has been tested over decades and whose judgment investors can underwrite. Anyone reaching for that framing — and the sector's promotional literature reaches for it constantly — is reaching for something that is not there.

Why This Is Section II and Not an Appendix

Almost every investment case for a young deep-technology company rests, at some level, on a claim about management: that these particular people can execute a plan that is objectively very hard. That claim has to be tested against the record. And the record here is that in roughly seven years, Iluvatar had a founder depart over a lost strategy fight, a chairman disappear into a corruption probe, and a third leader arrive from the finance side of the cap table.

Two chairman transitions and a founder exit in that span, one of them tied to a criminal investigation of a major state-linked group, is not a macro event that happened to the company. It is a data point about the company — about the stability of its strategic direction and the depth of its bench.

The company's own prospectus concedes the structural version of this. Unlike a typical founder-led business, ownership through the entire disclosed track record period has sat with employees holding stakes through special-purpose vehicles and with diversified passive financial investors. There has never been a single controlling shareholder.⁷ That is management's own language, in a legally binding document, and it is more candid than most of what has been written about the company since.

What it does not tell you is whether the thing they were arguing about back in 2016 — the general-purpose GPU — was worth all this. For that, you have to look at the silicon.

↑ Back to Top

III. Building the Product: Tiangai, Zhikai, and the CUDA-Compatibility Bet

Every Chinese GPU startup eventually walks into the same room and finds the same problem sitting in the chair: it is not the chip. It is the software.

Here is the simplest honest way to explain Nvidia's moat to someone who has never written a line of code. Imagine an industry where every factory in the world has spent fifteen years building custom tooling, training its workers, and writing its safety procedures around one manufacturer's machine. A competitor arrives with a machine that is genuinely as good — faster, even, on some jobs. It does not matter. Switching means retraining every worker, rewriting every procedure, and rebuilding every jig, and during that transition the factory produces nothing. The competitor is not fighting the machine. It is fighting fifteen years of accumulated habit.

Nvidia's machine-tooling is called CUDA. It is the software layer that lets a programmer tell a GPU what to do, and by the mid-2010s essentially every AI researcher on earth had learned to speak it. Frameworks like PyTorch — the software most AI models are actually written in — were built assuming CUDA underneath. That accumulated habit, not transistor count, is the deepest part of Nvidia's advantage, and it is the part that cannot be replicated by a foundry.

The Choice: Don't Fight the Habit, Inherit It

Iluvatar's central strategic decision was to refuse that fight.

Rather than ask customers to abandon CUDA and learn something new — the approach that has killed a long line of AI-chip startups worldwide — the company built for what it calls easy migration: deep compatibility with CUDA and with mainstream frameworks like PyTorch at the software interface layer, with its own proprietary stack underneath.³ Its compiler, its device drivers, and its libraries for mathematics and inter-chip communication — carrying names like ixDNN, ixBLAS, and ixCCL — do the real work, while presenting the programmer with something close enough to what they already know that porting an existing model becomes a project of days rather than quarters.³

In the factory analogy: keep every jig, every procedure, and every trained worker. Change the machine underneath and make it accept the same tooling.

This is a legitimately clever strategy and it deserves to be recognized as one, because it converts Nvidia's greatest strength into a shared standard that the challenger can free-ride on. It is also, for exactly the same reason, a strategy with a built-in ceiling. A company whose value proposition is "works like the incumbent's, but domestic" is competing on availability and price, not on a differentiated capability of its own. It inherits the ecosystem, but it does not own it, and it cannot shape where the standard goes next. If Nvidia's software layer evolves in a direction that is expensive to track, the compatibility burden becomes a treadmill that has to be run forever, funded out of an R&D budget an order of magnitude smaller than the incumbent's.

Two Chips, Two Jobs

The product line splits along the natural fault line of the AI workload.

天垓 Tiangai is the training family. Training is the brutally expensive part: you take a raw model and a mountain of data and you run enormous numbers of calculations for weeks to teach it anything at all. Training clusters are where the money and the prestige are, and where Nvidia's advantage is most complete.

智铠 Zhikai is the inference family. Inference is what happens afterward — the finished model answering a user's question, generating a paragraph, classifying an image. Any single inference is cheap. But a consumer AI product with tens of millions of users runs billions of them a day, and in aggregate inference is the larger and faster-growing compute market, with a crucial commercial property: it is far less sensitive to being the absolute fastest chip in the world. Inference buyers care about performance per dollar, per watt, and — in China specifically — per unit actually available for purchase.

Iluvatar was the first Chinese company to mass-produce a domestic GPGPU in each category, both on a 7-nanometre node.³² The dual-line structure matters strategically, not just as a product-catalogue fact. A training-only vendor in China is competing directly against the workload where Nvidia's residual advantage is largest and where Huawei's state-backed Ascend clusters have the deepest institutional relationships. An inference-capable vendor gets to compete in the segment where "good enough and actually purchasable" is a winning proposition.

The Foundry Ceiling

Now the constraint that governs everything else.

Iluvatar's advanced-node capacity comes from SMIC, domestically — not from TSMC.⁸ This is not a preference; it is the operating reality of the export-control environment that the entire Chinese GPU cohort lives inside.

The consequence is structural and cannot be engineered around. TSMC's leading-edge nodes are, by broad industry consensus, generations ahead of what is available to SMIC, and SMIC's advanced-node output is a scarce, nationally contested resource shared across every Chinese company that needs it — Huawei's chip operations, the four GPU startups, telecom silicon, and more. Nvidia designs a chip and buys as much leading-edge capacity as its cash and TSMC's roadmap allow. Iluvatar designs a chip and takes an allocation from a domestic foundry that is simultaneously behind on process and oversubscribed on volume.

So even if Iluvatar's designers were, chip-for-chip, as good as Nvidia's — a claim nobody has independently established — the company would still be manufacturing on a less advanced process with a hard limit on how many units it can produce. That is why shipment volume, not just design wins, is the number that reveals whether this business is real. And it is why any bullish scenario for the entire cohort is, at some level, a bet on SMIC's process and capacity roadmap rather than on any individual design team.

The Roadmap That Raised the Stakes

In January 2026, days after listing, Iluvatar did something unusual for a company with almost no public track record: it published a specific, dated, multi-generation product roadmap benchmarked against Nvidia by name.⁹¹⁰

The claims were bold. The company said its current Tianshu architecture already outperformed Nvidia's Hopper generation — the H200 class — on real DeepSeek V3 workloads, at over 90% utilization.⁹ It said a next architecture, Tianxuan, would target parity with Nvidia's Blackwell generation during 2026. And it said a further generation, Tianquan, would target Nvidia's Rubin platform by 2027.⁹¹⁰

Utilization is the key word in the first claim and deserves unpacking, because it is where domestic accelerators most often disappoint in practice. A chip's headline specification is its theoretical peak throughput. Utilization is how much of that peak a real workload actually achieves once the software stack, memory bandwidth, and inter-chip communication have taken their cut. Chinese accelerators have frequently posted competitive paper specifications and then delivered a fraction of them on real models, because the software layer was immature. A credible claim of high utilization on a demanding real-world model is therefore a more meaningful assertion than a raw specification comparison — if it holds up.

That is the problem: no independent verification of these benchmark claims has been found in public sources. They are self-reported, workload-specific, and selected by the company. A vendor benchmark showing a vendor's chip winning is not evidence of parity; it is evidence of what the vendor chose to measure.

What the roadmap does accomplish is to convert a vague promise into a testable one. Iluvatar has now publicly committed to specific competitive positions on specific dates against a competitor whose own release cadence is public. The specificity raises the stakes on execution far more than it demonstrates current parity.⁹ Investors get something valuable from that: a falsification schedule. Either Tianxuan silicon lands in 2026 and performs near where the company said it would, or it does not, and the market will find out on a known timetable.

Product claims are one thing. What customers actually paid is another, and that record is more mixed.

↑ Back to Top

IV. Scaling Revenue, Scaling Losses (2022–2025): The Segment and Unit-Economics Story

There is a chart that describes Iluvatar CoreX better than any narrative can, and it has two lines on it. One goes up steeply. The other goes down steeply. They are the same business.

Revenue climbed from RMB 189 million in 2022 to RMB 289 million in 2023, then RMB 540 million in 2024, then RMB 1.03 billion in 2025 — a compound growth rate of roughly 69% across the period, and 91.6% in 2025 alone.⁷¹¹ Very few businesses of any kind grow like that for four consecutive years.

Losses over the same period ran RMB 554 million, RMB 817 million, and RMB 892 million across 2022 through 2024 — cumulative losses of more than RMB 2.2 billion by the end of 2024 — with roughly RMB 1.0 billion more added in 2025.⁷¹² By the time the company reached the Hong Kong market, accumulated losses were comfortably north of RMB 3 billion.

Set those two sequences beside each other and the uncomfortable arithmetic becomes visible: through 2025, every incremental yuan of revenue growth arrived alongside an incremental yuan of loss. The business was not, on this evidence, scaling toward profitability. It was scaling.

What Was Actually Being Sold

The mix inside that revenue growth is more encouraging than the headline, and it is where the strategic bet showed up in the accounts.

GPGPU products alone accounted for RMB 923 million of 2025 revenue — 89.3% of the total — growing 149.6% year over year.¹¹ The fastest-growing piece was inference: the Zhikai line grew 238.2% in 2025.¹¹

That last number is the most informative single figure in the company's 2025 results, and it validates something real. The decision to build a dual-line business rather than chase training glory alone is showing up in the mix, not merely in the product catalogue. Inference is the segment where a domestic challenger's disadvantages matter least and where volume is structurally larger. A company whose inference line is growing more than three times as fast as its overall business is a company finding demand in the part of the market where it has the best structural chance of being genuinely competitive rather than merely available.

From 22 Customers to 340

The customer base tells a parallel story of de-risking. Iluvatar went from 22 customers in 2022 to more than 340 by the end of 2025, with over a thousand deployed projects spanning finance, healthcare, and transportation.³⁷

That breadth is meaningful in a sector where the default assumption — often correct — is that a domestic chip company's revenue is a handful of state-directed purchases dressed up as a market. Three hundred and forty customers across three unrelated industries is not that pattern. It suggests something closer to genuine, dispersed commercial adoption.

Two caveats keep this from being conclusive. Customer count is not customer concentration; a company can have 340 customers and still take most of its revenue from three of them, and the disclosed concentration percentages in the prospectus are the figures that settle it, not the headline count.⁷ And in China, "customer" spans a wide range of buyer types, from genuinely commercial enterprises to state-linked institutions purchasing under indigenization mandates. A broad base of policy-driven buyers is more durable than a narrow one, but it is still policy-driven demand.

The "Flour Costs More Than Bread" Problem

The structural reason this cohort loses money is not incompetence. It is a margin arithmetic that the Chinese trade press has captured in a phrase worth borrowing: 面粉贵过面包 — the flour costs more than the bread.⁸

The mechanics run as follows. Domestic GPU makers operate at gross margins around 50%, against Nvidia's roughly 75%.⁸ So on every yuan of revenue, an Iluvatar has about two-thirds as much gross profit as Nvidia does to fund everything else. And the largest thing it must fund is research and development, which in leading-edge chip design is a broadly fixed cost — designing a competitive GPGPU costs something close to the same whether you sell 40,000 units or four million.

Nvidia spreads that fixed cost across a global installed base measured in millions of units. Iluvatar spreads it across tens of thousands, from a smaller gross margin, while spending a similar or higher share of revenue on R&D just to stay on the treadmill. The result is the loop the Chinese press describes as 低毛利、高研发、高亏损 — low margin, high R&D, high loss.⁸

This is the single most important thing to understand about the entire domestic GPU cohort, and it is why the losses are not a management scandal. They are the physics of subscale semiconductor design. Which also means the losses cannot be fixed by cost discipline; they can only be fixed by volume. That is the mechanism, and it points directly at the metric that matters: the trajectory of units shipped and the price realized on them, not quarterly expense management.

One Business, No Hiding Places

A note on what this company is not, because it shapes how everything else should be read.

Iluvatar CoreX is a single-product-family business. GPGPU chips and the AI computing solutions built around them are essentially the whole enterprise.⁷ There is no diversified second segment quietly generating cash, no legacy division subsidizing the moonshot, no services annuity to smooth the cycle.

For an investor this cuts in both directions and both are worth naming. There is nowhere to hide: if the chips do not sell, nothing else is coming. But there is also nowhere for management to hide results. Conglomerate structures let weak units be buried inside strong ones. Here the number is the number.

Which is exactly what made the first half of 2026 so instructive — because that was the period when the number appeared to change sign.

↑ Back to Top

V. The 2026 Inflection: ByteDance, the Placement, and a "Profit" That Isn't One

Hong Kong's IPO market in late 2025 had rediscovered an appetite it had lost for years, and the appetite was specifically for Chinese semiconductors. When Iluvatar CoreX opened its books, the retail tranche was oversubscribed roughly 414 times.¹ That is not a valuation judgment. That is a queue.

The Listing

The offer priced at HK$144.60 per share, raising approximately HK$3.68 billion — about US$475 million.¹¹³ Eighteen cornerstone investors took 43% of the base offer, and the identity of two of them is worth pausing on.²

中兴通讯 ZTE, the telecom equipment giant, came in as a cornerstone — and ZTE was also a customer.² So did 第四范式 4Paradigm, an enterprise AI software company.² There is a legitimate reading of this: strategic investors with technical competence to evaluate the product put capital behind it, which is a stronger signal than a generalist fund doing the same. There is also a reading an activist would offer: when a customer becomes an anchor shareholder, the independence of that customer's purchasing decision becomes harder to assess from the outside, and revenue from a party that also owns your stock deserves more scrutiny than revenue from an arm's-length buyer. Both readings are available; neither is proven.

After the opening pop, the shares closed their first day up 11.1%.² Then the run began in earnest. By May 2026 the stock had nearly tripled from the offer price. By early July it was up roughly 234% since listing.⁴¹⁴

At which point the company did two things with the market's enthusiasm.

Raising While the Window Was Open

On July 13, 2026 — six months and five days after listing — Iluvatar placed HK$7.07 billion of new H-shares, approximately US$902 million, at a 15% discount to the prior close.¹⁵ It ranked among the largest secondary raises by a Hong Kong-listed semiconductor company that year.

Combined with the IPO, that is more than HK$10.7 billion raised in roughly six months by a company that had never generated operating profit.

The 15% discount is the detail worth dwelling on, because it is the price of speed and size. A placement at a steep discount transfers value from existing holders to the new buyers, and it is what a company accepts when it wants a very large amount of money quickly rather than the best possible price. Doing it into a stock that had roughly tripled is defensible capital-markets practice — raise when you can, not when you must, and a company burning cash on leading-edge chip development absolutely should build a war chest when the window is open. But it is worth being precise about what it demonstrates. It demonstrates competent opportunism about the equity window. It does not demonstrate that the underlying business is funding itself, and it establishes a pattern of financing that dilutes rather than compounds.

Then, on August 20, 2026, the board proposed a capitalization issue: two new shares for every one held, expanding the share count roughly threefold from about 269 million to about 808 million shares.¹⁶ No new capital was raised. No shareholder's proportional stake changed. The entire purpose was to bring the per-share price down and improve trading liquidity — a cosmetic act, correctly understood, though not a meaningless one for a high-priced stock with heavy retail participation.

The ByteDance Order

In June 2026, reports emerged that carried more genuine information about this company's prospects than every benchmark claim it had ever published.

字节跳动 ByteDance was reported to be negotiating the purchase of at least 50,000 inference chips from Iluvatar for 豆包 Doubao, its consumer AI assistant.¹⁷ Guosen Securities estimated the order's value at close to RMB 600 million — roughly 60% of the company's entire prior-year revenue, from one customer, in one deal.¹⁷ It would also exceed the company's total GPU shipment volume for all of 2025.

Why this matters more than a benchmark: ByteDance is not a state institution fulfilling an indigenization quota. It is the operator of one of the most demanding consumer AI workloads in China, with the engineering depth to evaluate silicon rigorously and no obligation to buy domestic if domestic does not work. A company like that qualifying Iluvatar chips for production inference — not a pilot, not a proof of concept, but the volume that serves real users — is the most credible external validation available in this industry. Such evidence has been remarkably thin across the entire domestic cohort, which is precisely why one order carries this much weight.

Two disciplines apply. First, a negotiation reported in the press is not a signed contract, and a signed contract is not booked revenue, and booked revenue is not collected cash. Each of those conversions has to actually happen. Second — and this is the harder point — a single order equal to 60% of prior-year revenue is simultaneously the best news in the story and a concentration risk of the first order. A customer that large sets the price. If ByteDance repeats in 2027, Iluvatar has a franchise. If it does not, the company has had one very good year and a hole in the next one.

The Profit That Wasn't

Now the passage that matters most in this entire article, and it sits deliberately right here, next to the good news, rather than in a risk section thirty paragraphs later.

For the first half of 2026, Iluvatar reported revenue of approximately RMB 946 million, up 191.6% year over year, and a net profit of roughly RMB 106 million — its first profitable period in company history.¹⁸⁴ The headlines wrote themselves: the Chinese GPU maker had turned the corner.

It had not.

In April 2026, Iluvatar made a RMB 100 million strategic investment in 盛合晶微 Shenghe Jingwei, a chip-testing company that subsequently listed on the STAR Market.⁴ The listing went well. Iluvatar's stake was marked to the post-IPO price, generating an unrealized fair-value gain of approximately RMB 756 million.⁴

That gain is where the profit came from.

The shares are locked up for twelve months. They cannot be sold. No cash entered the business. The gain is an accounting recognition of a paper value on a holding the company is contractually barred from monetizing — and one that will be re-marked, up or down, at every subsequent reporting date depending on where an unrelated company's stock price happens to sit.

Strip it out and the picture inverts. On a like-for-like operating basis, the core GPU business lost roughly RMB 654 million in the first half of 2026 alone — a wider core loss in six months than the company had recorded in any full prior year.⁴

Sit with the two facts together. Revenue nearly tripled. The core loss widened to a record. That is not a turnaround narrative interrupted by an accounting quirk; that is an accounting quirk obscuring the absence of a turnaround.

What This Says About Management

Two claims that any investor might reasonably want to make about this management team run straight into this disclosure, and both need to be revised rather than dismissed.

The first is the capital-allocation claim. Management has demonstrably been effective on the capital-markets side: a well-received IPO, an opportunistic mega-placement, and a strategic minority investment that produced a large paper gain within months. That is a real skill and it is not trivial — the balance sheet it produced buys years of runway that many peers do not have. But note what all three of those wins have in common. They are financial transactions. Not one of them is evidence of converting chip revenue into chip profit, which is the thing the company exists to do.

The second is the turning-the-corner claim, and the company's own interim disclosure is the strongest disconfirming evidence available against it. This is not an outside analyst's skeptical modelling. It is Iluvatar's own numbers, in its own filing, showing that operating losses widened even as revenue nearly tripled.¹⁸⁴

Where does that leave the claim? Not rejected outright — revenue growth of 191.6% is real demand, and a business scaling this fast will naturally run negative operating leverage for a period while it invests ahead of volume. But the claim has to be narrowed sharply, from "the company has turned profitable" to "the company has not yet demonstrated operating leverage, and the direction of travel in the most recent period was the wrong one."

And the test that would resolve it is specific and near-term: does core-business operating loss — excluding all investment and fair-value gains — narrow in the second half of 2026 and in the full-year results, and does the ByteDance volume convert into booked, cash-collected revenue? Those two questions, answered over the next two reporting periods, settle more than any roadmap slide can.

To understand whether they can be answered favourably, the next thing to examine is the field Iluvatar is playing on — because most of the forces acting on this company act on all of its competitors too.

↑ Back to Top

VI. Industry Structure: The "Four Little Dragons," Huawei, and Nvidia

In the space of roughly a year spanning late 2025 and 2026, four Chinese GPU companies — none of them profitable, all of them subscale, all selling into the same protected market — either listed in Hong Kong or filed to.⁸ The Chinese press gave the group a name with the faint irony such names usually carry: 国产GPU四小龙, the four little dragons of domestic GPUs.

摩尔线程 Moore Threads. 沐曦股份 MetaX. 壁仞科技 Biren Technology. And Iluvatar CoreX.⁸

Four companies, one thesis, one customer pool, one foundry, one policy tailwind, and one very obvious question: what happens when the tailwind is shared by everyone?

The Tailwind Is Enormous and It Is Real

Start with the demand, because it is not imaginary. China's GPGPU market grew at a 72.8% compound annual rate from 2022 to 2024, reaching roughly 1.6 million units, with domestic share projected to exceed 50% by 2029.²⁸

The driver is straightforward. US export controls progressively restricted Nvidia's most capable accelerators from the Chinese market. Chinese demand for AI compute did not decline in response — if anything it accelerated, as domestic model developers raced to keep pace. A large, urgent, well-funded demand pool was severed from its preferred supplier. Somebody was going to fill it.

That is the single biggest reason every company in this cohort commands the valuation it does, and it is worth stating plainly that this tailwind does not require any of them to beat Nvidia on merit. It only requires them to be available.

But here is what the same fact means when read the other way: the protection accrues to the category, not to any company within it. Import substitution creates a pool of demand that must be served domestically. It does not decide which domestic supplier serves it, and it does nothing to fix the subscale economics that afflict all four dragons equally. A policy that guarantees the industry's revenue while guaranteeing nobody's profit is a strange thing to underwrite as a moat.

And Nvidia has not actually left. Even under restriction, it holds an estimated 70% of China's AI accelerator market, through H20-class compliant parts and gray-market channels.⁸ The four dragons plus Huawei plus Cambricon are contesting the remainder. The domestic-substitution story is, at this stage, a fight over the minority of the market — with the constant background risk, cutting the other way, that any easing of restrictions returns the incumbent to full strength at exactly the moment domestic vendors have committed capital assuming it would not.

Where Iluvatar Actually Ranks

The comparison across the cohort produces a genuinely interesting split.

On 2024 shipment volume, Iluvatar led at 38,000 units, against MetaX at 24,000, Cambricon at 26,000, and Enflame at 13,000.⁸ The company likes this framing, and it is not wrong to like it — units shipped is the number closest to real deployment.

On 2024 revenue, Iluvatar trailed. MetaX booked RMB 743 million, 燧原科技 Enflame RMB 722 million, Moore Threads RMB 438 million, and Biren RMB 337 million, against Iluvatar's RMB 540 million.⁸

Shipping the most units while booking less revenue than two peers means one thing: the average selling price is lower, and the product mix is weighted toward cheaper parts. That is not a scandal — it may be a deliberate strategy of buying installed base and developer familiarity at low prices, which for a company whose entire moat thesis rests on ecosystem adoption could be rational. But it does mean that "shipment leader" and "market leader" are not the same claim, and the company's promotional materials tend to blur them.

The blur becomes untenable at the next level of detail. Disclosed 2024 domestic GPGPU market share for Iluvatar was under 1% — by some estimates around 0.3%.⁸ Every use of the word "leader" in this company's materials needs to be read against that number. It is a leader within a cohort of loss-making startups that collectively hold a small slice of their own domestic market.

The One Comparison Iluvatar Wins Cleanly

There is, however, a peer comparison where Iluvatar's numbers look genuinely and importantly better.

In 2025, Iluvatar lost roughly RMB 1.0 billion on revenue of about RMB 1.03 billion. Biren lost approximately RMB 16.5 billion on similar revenue.¹²

Read that again, because the ratio is extraordinary. Two companies, essentially the same top line, with loss figures separated by more than an order of magnitude. Whatever accounting and one-off components sit inside Biren's figure, the gap is far too large to be explained away, and it says something real about the operating discipline of an organization that has been run, since 2023, by someone from the audit and private-equity world rather than by a technologist.

This is the strongest single argument in favour of Iluvatar's management, and it is the one place where the non-engineer CEO looks like a feature rather than a gap. In a category defined by capital incineration, burning an order of magnitude less than a direct peer to reach the same revenue is a competitive advantage in itself — it extends runway and reduces the frequency of dilutive raises. It is also, notably, a strength that does not depend on any unverified benchmark claim.

The Competitors Who Aren't in the IPO Queue

Two companies loom over this whole cohort without being part of it.

华为 Huawei's 昇腾 Ascend line — the 910C, positioned against Nvidia's H20, and the newer Ascend 920 — sits entirely outside the four dragons' world.⁸ Huawei is privately held, state-backed at a level no startup can match, and vertically integrated across chip, system, framework, and cloud in a way that structurally advantages it in enterprise and government accounts. On raw at-scale credibility as a domestic Nvidia alternative, Ascend is ahead of everyone in this story, even if it remains behind Nvidia's high end.

That is a substitution threat of a particular and dangerous kind. Huawei is not merely another chip vendor; it can bundle silicon into a full stack that a customer buys as one decision. A startup selling a component into an account where Huawei sells the whole system is in a structurally weaker position regardless of how good the component is.

寒武纪 Cambricon is the other shadow, and a different kind of one. It is the domestic AI-silicon company that has actually reached profitability, trading at rich multiples on the strength of it.⁸ Cambricon is the bar. Every one of the four dragons is being valued, implicitly, on the proposition that it will eventually look like Cambricon. None of them has yet shown it can.

Porter's Five Forces, Applied Honestly

Run the framework and the picture is sobering.

Buyer power is high and rising. Hyperscalers like ByteDance can multi-source and do, and Beijing's own policy preference for vendor diversification structurally strengthens buyers against any single domestic supplier. The state does not want one national champion in GPUs; it wants several, which is excellent industrial policy and poor news for the pricing power of each.

Supplier power is high. SMIC's advanced-node capacity is scarce and contested across the entire Chinese electronics industry. When your only viable foundry is oversubscribed and your competitors are in the same queue, allocation becomes a political and relationship question rather than a commercial one.

Threat of substitution is significant, and it comes from Huawei's vertically integrated stack more than from any of the dragons.

Rivalry is intense and margin-compressing. Four subscale competitors chasing the same customers with broadly similar propositions, all newly flush with IPO capital and all needing to demonstrate volume growth to their new shareholders, is a recipe for price competition in a category that already cannot afford it.

Threat of new entrants is moderate — capital requirements and design talent scarcity are genuine barriers — but the four dragons themselves are evidence that the barrier was crossable within a decade, and China's capital markets have shown enthusiasm for funding the next attempt.

Where the Moat Claim Actually Stands

The CUDA-compatibility strategy is a real technical choice, verifiable in the company's product documentation and roadmap disclosures.³⁹ The question is whether it has produced a moat, and the honest answer is that there is not yet enough evidence in either direction.

A moat in semiconductors is tested through three mechanisms: whether customers multi-source, whether they stay, and whether the vendor can raise price. On multi-sourcing, the answer is plainly yes — every serious Chinese AI buyer runs multiple accelerator vendors deliberately, as both engineering insurance and negotiating leverage. On retention, the public record is thin: a customer base that expanded from 22 to 340 in three years has not existed long enough at scale to produce a meaningful churn or repeat-purchase history. On price realization, the ASP evidence points the wrong way — the shipment-leader-but-revenue-laggard pattern is what a company without pricing power looks like.

So the calibrated conclusion: the moat claim is not rejected, but it is unproven and currently unsupported by the price evidence. It should be held at the confidence its affirmative evidence supports, which is low. The ByteDance order is the first genuine data point that a sophisticated hyperscaler will place Iluvatar silicon into production alongside its other suppliers — which narrows, without resolving, the question of whether this is a real business or a subsidised also-ran.

And the entity that decides how that question gets answered is the group of people running the company — which is where the picture becomes less comfortable again.

↑ Back to Top

VII. Management, Ownership, and Incentives

There is a question worth asking of any newly public company, and it is not "who is the CEO." It is "who actually decides." At Iluvatar CoreX, that question has an unusually elaborate answer.

The People

盖鲁江 Gai Lujiang holds both chairman and CEO titles. His background is audit and private equity — PwC, Deloitte, then a Zhongzhi-linked investment platform — and he arrived at the top of the company as an investor-side board representative after the second leadership vacancy, not as a technologist rising through the engineering organization.⁵

The combination of chairman and CEO in one person is itself a governance point. It concentrates board oversight and executive authority in the same hands at a company that has already demonstrated it lacks a stable counterweight — no controlling founder, no dominant strategic shareholder to push back.

The technical anchor sits elsewhere. Chief Scientist 郑金山 Zheng Jinshan previously led GPU design work for AMD's Shanghai team, which is a genuinely relevant pedigree: AMD is one of very few organizations on earth with real institutional knowledge of building a competitive general-purpose GPU against Nvidia.⁷

That division of labour — finance at the top, engineering below — is worth tracking as its own risk. In a deep-technology company, the departure of the chief scientist is a more meaningful negative signal than the departure of most executives, because the technical judgment on which the entire roadmap rests is concentrated in a small number of people. Zheng's continued presence is a variable investors should watch on its own, independent of the CEO.

Who Owns It

There is no controlling shareholder. That fact, disclosed plainly in the prospectus, structures everything else.⁷

The largest single bloc is 上海数麒 Shanghai Shuqi, the general partner representing several employee and management special-purpose vehicles, holding roughly 23.6%.⁷ Critically, that bloc does not vote as a diffuse collection of employees. It votes as directed by a management committee dominated by the company's own executive directors and senior management.⁷

The largest outside institutional holder was 大钲资本 DCP Capital at roughly 22.9% pre-IPO — a private-equity firm, not a strategic industrial partner.⁷

Set those two facts side by side and the governance reality becomes clear. This is a company whose largest voting bloc is controlled by a committee of the executives it is meant to oversee, with the second-largest position held by a financial sponsor whose eventual objective is a return, not a permanent industrial position. It is a founder-less, financial-sponsor-and-management-committee-governed company, and that should be named plainly rather than glossed, because it forecloses one of the most common framings applied to young technology companies: the aligned founder-operator with skin in the game and a multi-decade horizon. That framing is not available here.

Employee ownership through SPVs does create genuine alignment of a kind — the people building the chips benefit if the chips sell. But an employee bloc voted by a management committee is a different instrument from an employee bloc voting individually. It concentrates control at the top of the organization while distributing the economics beneath it.

The Capital Allocation Record, As Short As It Is

Here is a discipline worth applying rigorously: what has this management team actually done with capital since it acquired the ability to raise it publicly?

Two headline acts. A large secondary placement at a 15% discount seven months after listing. And a RMB 100 million strategic minority stake in another company's IPO that became the accounting source of the firm's only reported profit.

Neither is a red flag on its own. Raising into strength is defensible. A strategic stake in a chip-testing company sitting adjacent to its own supply chain has plausible industrial logic beyond the financial return.

But together they characterise a management team that has so far demonstrated more capability on the capital-markets side of the business than on the operating side. That is the plain reading of an eight-month public record, and it should be held lightly — eight months is not a track record, and a company that has just raised HK$10.7 billion has genuinely reduced its financing risk, which is worth something real. What it should not do is permit the inference that good capital allocation has been demonstrated. The stock's post-IPO performance is not evidence of that. It is evidence of what buyers were willing to pay.

Any activist looking at this company would go straight to three places: the combined chairman-CEO role at a business with no controlling shareholder; the management-committee-directed voting of the largest employee bloc; and the fact that the only reported profit in company history came from a mark-to-market gain on an illiquid, locked-up minority stake rather than from operations. None of those is evidence of wrongdoing. All three are structural features that reduce external accountability at precisely the moment external accountability matters most — the first years of public life, when habits form.

Guidance and Credibility: Too Early to Score, But Not Too Early to Watch

Assessing management credibility properly requires observing promises against outcomes across multiple cycles. Iluvatar has been public for eight months. That record does not exist yet, and pretending otherwise would be the analytical error here.

What does exist is one instructive early signal. The FY2025 results release in March 2026 reported that adjusted net loss narrowed 32.1% year over year — while the reported, unadjusted net loss widened.¹⁹¹¹

Both statements can be simultaneously true and both can be found in the same announcement. But the choice of which one leads the narrative is a management decision, and it is exactly the kind of decision that, repeated over several years, establishes whether a company communicates candidly or promotionally. A single instance proves nothing; a pattern would. What a careful reader does with it now is simple: note the adjustments disclosed, note which figure the company chose to headline, and watch whether the same framing recurs when the adjustments are less flattering.

The same discipline applies to the January 2026 roadmap. Management put dated, competitor-benchmarked commitments into the public domain within days of listing. That is either admirable specificity or a hostage to fortune, and the difference will be settled by delivered silicon rather than by argument.

With the people, the ownership, and the economics all on the table, the two competing cases can now be built properly.

↑ Back to Top

VIII. Bull Case vs. Bear Case

Two competent investors can look at the same facts about Iluvatar CoreX and reach opposite conclusions, and both arguments deserve to be made at full strength rather than as strawmen.

The Bull Case

The first pillar is that the technical firsts are genuine. Being the first Chinese company to mass-produce both a training and an inference GPGPU on a 7-nanometre node is a verifiable engineering achievement, not marketing.³² Chip design at this level is unforgiving — the design either yields working silicon in volume or it does not — and Iluvatar's cohort-leading 2024 shipment volume of 38,000 units is corroboration that working silicon reached customers at scale.⁸

The second pillar is that the roadmap engages seriously with reality. Committing publicly to specific competitive positions against named Nvidia generations on dated timelines is what a company does when it believes it has line of sight on execution. It is not proof, but it is a materially different posture from vague claims about world-leading performance.⁹¹⁰

The third pillar is the ByteDance order, and in the bull case it is the load-bearing one. Production-scale qualification by a commercially motivated hyperscaler with genuine engineering capability is exactly the evidence that has been missing from the domestic GPU story, and its arrival changes the character of the argument.¹⁷

The fourth pillar is loss discipline. The comparison with Biren — the same revenue, an order of magnitude less loss — is the rare case where a domestic GPU company's financial statements look better than a direct peer's rather than merely less bad than expected.¹² Runway is a competitive weapon in a business this capital-hungry, and Iluvatar is consuming it far more slowly than at least one rival.

The fifth pillar is the balance sheet. More than HK$10.7 billion raised in six months buys years of development regardless of near-term chip profitability, and it means the company is not dependent on the equity window staying open in order to fund its next generation.¹³¹⁵

The sixth pillar is the market itself. A 72.8% compound growth rate with domestic share projected past 50% by 2029 is a tailwind that does not require Iluvatar to beat Nvidia on merit.²⁸ It only requires the company to keep taking a growing slice of a protected and rapidly expanding pool — a far lower bar than technological superiority.

The Bear Case

The first count is governance, and it is not a footnote here. A founder who left over a lost strategy fight, a chairman who disappeared into a corruption investigation, a combined chairman-CEO from the finance side, and no controlling shareholder — replaced by a management committee that votes the largest employee bloc.⁷⁵ For a company whose investment case depends entirely on execution over many years, that is an unusually thin foundation of demonstrated stability.

The second count is that the profitability headline does not describe chip economics, and the core loss widened in the most recent period even as revenue nearly tripled.⁴ This is the company's own disclosure working against the company's own narrative.

The third count is structural margin disadvantage. Operating at roughly 50% gross margin against Nvidia's 75%, while carrying comparable fixed R&D, means the gap has to be bridged by capital raises rather than by operating cash flow.⁸ Every year that break-even recedes is another year of potential dilution — and the July placement established that management will accept a meaningful discount to get size done.

The fourth count is that sub-1% domestic market share sits awkwardly beneath the "leader" language, and the market it holds that share in is squeezed from both ends.⁸ Huawei's vertically integrated Ascend stack compresses it from the domestic side. Any easing of US restrictions would let Nvidia compress it from the other.

The fifth count is the foundry ceiling. Volume is bounded by SMIC's allocation and process, and no amount of design excellence or customer demand relaxes that constraint.

The sixth count is that the performance claims remain unverified by any independent party.

The seventh count is valuation, and it may be the most immediately consequential. Following the post-IPO run, the shares traded around a price-to-sales multiple in the region of 115 times, against a Hong Kong semiconductor sector average nearer 50 times.¹⁴ A multiple like that does not merely reflect optimism about the future; it capitalises years of successful execution as though it had already occurred. It leaves essentially no room for a disappointing print.

7 Powers: Which Ones Does Iluvatar Actually Have?

Hamilton Helmer's framework asks which of seven durable advantages a business possesses. Applied here, the results are thin, and honestly so.

Scale economies — no. Iluvatar is the subscale party in every relevant comparison; the fixed-cost-amortisation logic runs against it, not for it.

Network economies — partially, and indirectly. A developer ecosystem is a network, but Iluvatar's strategy is to plug into CUDA's network rather than build its own. It rents the network effect. It does not own it, and a rented network effect can be withdrawn or made expensive by the landlord.

Counter-positioning — arguably the strongest claim available. Iluvatar occupies a position Nvidia structurally cannot: a domestic Chinese supplier not subject to export restriction. Nvidia cannot counter this without a change in US policy, which is outside its control. But note that this is counter-positioning against a regulation, not against a business model, and it is shared identically with every other domestic vendor — which is the definition of a non-differentiating advantage.

Switching costs — unproven. The CUDA-compatibility strategy is explicitly designed to lower switching costs, which is precisely what makes migration to Iluvatar easy. The same property makes migration away from Iluvatar easy. This is a genuine strategic tension at the heart of the product design, and it has not yet been resolved by evidence either way.

Branding — no meaningful evidence. Buyers of AI accelerators evaluate on benchmarks, availability, and price.

Cornered resource — no, and the reverse is closer to true. The cornered resource in this industry is SMIC's advanced-node capacity, and Iluvatar competes for it rather than controlling it. Being an early domestic entrant confers relationships, not entitlement.

Process power — not demonstrated. Eight months of public life and a decade of losses do not evidence a superior, hard-to-replicate operating capability, though the loss-discipline comparison with Biren is the beginning of a case that something in this organization's operating rhythm is better than a peer's.

The scorecard is one plausible power (counter-positioning, shared with all competitors), one rented power (network economies), one contested power (switching costs, working in both directions), and four absent. That is a young company with a policy shelter, not a company with a moat — and it should be underwritten as the former.

Weighing the Two

The two cases are not symmetrical in the kind of evidence they rest on.

The bull case rests substantially on things that are true about the market: it is growing at extraordinary rates, it is protected, and domestic share is mandated to rise. Those things are well-evidenced and largely outside the company's control. Where the bull case depends on things about this company specifically — roadmap delivery, ByteDance conversion, eventual operating leverage — the evidence is either forward-looking or unverified.

The bear case rests substantially on things that are already in the historical record: the governance sequence, the margin structure, the market share, the widening core loss, the foundry dependency. Those are facts, not forecasts.

That asymmetry does not make the bear case correct. Growth companies are supposed to have their best evidence in the future; that is what growth means. But it does define what a position in this company actually is: an underwriting of a market and a policy environment, plus a bet that a management team with a short and financially oriented track record can convert a shared tailwind into company-specific economics that no member of its peer group has yet demonstrated.

The fairest characterisation is the one the evidence supports and no more: credible technical progress and a real hyperscaler win, funded by capital markets rather than by chip profit, run by a management team with an unusually thin operating and governance track record. That is neither a rejected thesis nor a proven one. It is an open one, with a defined set of tests coming.

↑ Back to Top

IX. What to Watch: KPIs, Catalysts, and Risks

Most companies present investors with a dashboard of a dozen metrics, of which two matter. Iluvatar is unusually clean in this respect, because the argument is so tightly defined.

The Numbers That Settle the Argument

First: core-business operating loss, excluding all investment and fair-value gains. This is the single number that confirms or falsifies the turning-the-corner story. Reported net profit at this company is now contaminated by the mark on a locked-up equity stake that will be re-valued at every reporting date according to an unrelated company's share price. Any reader who tracks reported net income is tracking the STAR Market as much as the GPU business. The clean figure — revenue minus the costs of actually designing, making, and selling chips — is the only one that reveals whether growth is producing operating leverage. It should be recomputed from the disclosure at every result, not taken from the headline.

Second: shipment volume and the direction of average selling price. Revenue growth alone cannot distinguish between two entirely different companies. One of them is winning volume while holding price, which is what a differentiated product looks like. The other is buying volume with price, which is what a commodity supplier does under competitive pressure. Given that Iluvatar already ships more units than peers who book more revenue, the ASP trend is the tell. A rising ASP alongside rising units would be the first hard evidence of pricing power. A falling one, in a market growing this fast, would say the competitive dynamic among the four dragons has already turned against everybody.

Third: whether the ByteDance relationship becomes a repeat relationship. Watch it through three gates in sequence — reported negotiation to signed order, signed order to booked and cash-collected revenue in the FY2026 accounts, and one-off order to recurring purchases in 2027. Passing the first two gates would validate the product. Only the third would begin to demonstrate that customer retention exists at hyperscaler scale, which is the moat question the current record cannot answer.

Catalysts on the Calendar

Full-year 2026 results, expected around March 2027, are the first meaningful test of all three metrics together, and the first opportunity to see whether the second-half operating trajectory diverged from the first.

Any news on SMIC capacity allocation matters more to this company's volume ceiling than most product announcements do, and it arrives on nobody's predictable schedule.

Roadmap delivery is the visible test of credibility: Tianxuan silicon and the Blackwell-parity claim were committed to 2026, and that year is now three months from closing.⁹

And on the technical side of the share register, further placements or lock-up expiries deserve attention. The cornerstone lock-up was already tested in July 2026, and a company that has raised twice in six months has established that it will use the equity market when it is open.¹⁵ Supply from either source can pressure the stock independently of anything happening in the business.

The Risks That Would Actually Break It

Execution risk sits first, because the roadmap is aggressive and the manufacturing constraint is real. Committing publicly to matching Nvidia's generational cadence while sourcing from a foundry that is behind on process is a bet on design efficiency overcoming a manufacturing deficit. It may work. It is not the way to bet if the constraint tightens.

Customer-concentration risk arrives with success, which is the awkward part. An order equal to roughly 60% of prior-year revenue makes the customer, not the supplier, the party with leverage in the next negotiation. Winning ByteDance and then discovering what ByteDance's bargaining power feels like are the same event, separated by a year.

Geopolitical and export-control risk runs in both directions simultaneously, and this is frequently misunderstood. Tighter US restrictions on Nvidia sustain the domestic tailwind that makes this whole cohort investable. But Iluvatar's own supply chain — SMIC capacity, electronic design automation tools, licensed IP blocks — is itself exposed to further restriction. The policy that creates the opportunity is administered by the same authority that could constrain the means of pursuing it.

Financing and dilution risk follows directly from the margin structure. A company that funds development from the equity market rather than from operations is exposed to that market's mood. The current balance sheet makes this a multi-year rather than a near-term concern, but the mechanism has not been eliminated — only deferred.

Governance risk is the pattern already established rather than any single event: leadership turnover without a stable successor bench, no controlling shareholder, and a combined chairman-CEO whose expertise is financial rather than technical, at a company whose central challenge is technical.

And valuation risk is simply the arithmetic of a very high multiple meeting a company with a narrow proof base. At a P/S multiple more than twice the sector average, the market has already paid for several years of the bull case.¹⁴ The consequence is asymmetry: good news is substantially in the price, while a disappointing print on any of the three KPIs above has considerably more room to move the stock than a reassuring one does.

↑ Back to Top

X. Durable Lessons & Closing

Some companies are worth studying because of what they are. Others are worth studying because of what they illustrate. Iluvatar CoreX, eight months into public life, is mostly the second kind — and it illustrates three things unusually cleanly.

Protection Is Not Profit

The first lesson concerns investing in a policy-protected industry. Government-driven import substitution can manufacture years of revenue growth, and the growth is entirely real — a 72.8% compound rate in China's GPGPU market is not an accounting artifact.⁸ But protection operates at the level of the category. It creates a pool of demand that must be filled domestically, and then it stands back while four subscale companies with similar products and shared cost structures compete to fill it.

The shelter is real. What it shelters is the whole loss-making cohort together, not any individual company's economics. An investor who mistakes a category tailwind for a company moat has confused the reason the industry exists with the reason one firm within it might eventually earn a return. The two are separate questions, and only the second one is answerable from company-specific evidence.

Read the Word "Profit" Before Believing It

The second lesson is more portable, and it applies far beyond Chinese semiconductors.

When a company that has never been profitable announces that it has become profitable, the correct first response is not celebration but decomposition. Where did the number come from? Is it operating? Is it cash? Is it recurring? Is it even realizable?

In this case the answer to all four questions was no. A RMB 100 million investment made in April 2026 produced a paper gain of roughly RMB 756 million by mid-year, and that gain — on shares that could not be sold — was the entire reported profit and then some.⁴ The underlying business had its worst six months on record.

Nothing about the accounting was improper. Fair-value marks on equity holdings are standard practice. The failure, where it occurred, was in the reading. Investment gains, fair-value marks, and one-off items all have to come out before an operating turnaround can be claimed, and the discipline of doing that arithmetic yourself is the difference between owning a business and owning a headline.

Governance Has to Be Built, Not Assumed

The third lesson concerns newly public, founder-less, sponsor-governed companies, of which there will be many more as this listing wave continues.

The standard shortcuts do not work here. There is no founder whose long-term alignment can be assumed from the size of a personal stake. There is no controlling shareholder whose interests can be read off a cap table. There is a management committee voting an employee bloc, a private-equity firm with a return horizon, and an executive who came from the audit profession.

None of that is disqualifying. It simply means that every judgment about capital allocation, strategic consistency, and candour in communication has to be built from evidence accumulated after the listing, and cannot be inferred from structure beforehand. That construction takes years, and it begins with the small choices — like which of two true loss figures a company puts at the top of its results announcement.

Where This Leaves the Story

What Iluvatar CoreX carries into 2027 is a genuinely mixed inventory. Real engineering firsts. A real hyperscaler order that no domestic competitor can match as evidence. A balance sheet strong enough to survive several years of being wrong. And, set against all of that, an operating business whose losses were widening at the last report, a market share still under one percent, a foundry it does not control, a management team it has not had long, and a share price that has already been paid for a future the company has yet to deliver.

Whether the technology becomes a business — that is, whether working chips shipping in volume ever become chips shipping at a profit — is not a question the last two years of stock performance can answer. It is a question that gets answered in the next two or three reporting periods, in the unglamorous arithmetic of core operating loss, average selling price, and whether the customer who placed the biggest order in company history decides to place another one.

References

1. Chinese GPU-Maker Iluvatar CoreX Seeks $475 Million in Hong Kong Listing — Caixin Global, 2025-12-30 ↩↩↩

2. Iluvatar CoreX's Hong Kong IPO: Hardcore Breakthrough Battle of China's General-Purpose GPU "Leader" — ACN Newswire via The Globe and Mail, 2025-12-31 ↩↩↩↩↩↩↩↩↩↩↩

3. Shanghai Iluvatar CoreX Semiconductor Co., Ltd. — Official Website / Investors ↩↩↩↩↩↩↩↩

4. 天数智芯扭亏背后：营收增191%，超7亿股权投资浮盈托底利润 — 时代周报 via Investing.com, 2026 ↩↩↩↩↩↩↩↩↩

5. 创始人出走、两次换帅、战略反复摇摆，天数智芯终于要上市了 — 新浪财经, 2026-01-05 ↩↩↩↩↩↩↩↩

6. 创始人出走、两次换帅、战略反复摇摆，天数智芯终于要上市了 — 腾讯新闻, 2026-01-05 ↩↩

7. Prospectus / IPO Offering Circular (Chinese) — HKEXnews, 2025-12-30 ↩↩↩↩↩↩↩↩↩↩↩↩↩

8. 四小龙齐聚：天数智芯5万出货难掩"面粉贵过面包"的焦虑 — 腾讯新闻, 2025-12-29 ↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩

9. China's Iluvatar CoreX unveils four-generation GPU roadmap aimed at surpassing Nvidia Rubin — Tom's Hardware, 2026-01 ↩↩↩↩↩↩↩

10. Iluvatar CoreX lays out GPU roadmap to surpass Nvidia by 2027 — KrASIA ↩↩↩

11. 天数智芯2025年营收同比增长91.6%，产品持续迭代升级 — 东方财富网, 2026-03-31 ↩↩↩↩

12. Biren, Iluvatar CoreX post triple-digit revenue growth but losses persist in AI chip race — South China Morning Post ↩↩↩

13. Kirkland Advises Iluvatar CoreX on its Hong Kong IPO — Kirkland & Ellis, 2026-01 ↩↩

14. 天数智芯 (09903) 股票股价、公司资料 — 雪球 (Xueqiu) ↩↩↩

15. Kirkland Represents Iluvatar CoreX on its HK$7.07 Billion H-Share Placing — Kirkland & Ellis, 2026-07 ↩↩↩

16. 天数智芯 (09903) 最新公告和新闻稿 — 富途牛牛 (Futu) ↩

17. 5万GPU芯片大采购背后：国产巨头天数智芯浮出水面 — 21世纪经济报道, 2026-06-24 ↩↩↩

18. 2026 Interim Results News Release — HKEX, 2026-08-19 ↩↩

19. 2025 Annual Results Announcement — HKEXnews, 2026-03-30 ↩