Shanghai Iluvatar CoreX Semiconductor (9903.HK) - Deep Dive Research Report

Prepared 2026-06-14. Currency: figures in RMB unless stated; IPO and market figures in HKD/USD as noted.

A note on sources and reporting cadence

Iluvatar CoreX (Chinese name 上海天数智芯半导体, "Tianshu Zhixin") listed on the Hong Kong main board on 8 January 2026. It reports on a half-yearly cadence and has published exactly one results set as a listed company: its FY2025 annual results on ~1 April 2026. There are therefore no five quarterly earnings-call transcripts to draw on - the company has been public for roughly five months. Where the report calls for "concall" material, I use the five disclosed reporting periods in the IPO prospectus and post-IPO disclosure (FY2022, FY2023, FY2024, H1 2025, FY2025) plus the company's 26 January 2026 architecture-roadmap investor event, and I flag this limitation explicitly in Sections 7 and 9 rather than inventing call transcripts.

Section 1: What the company does

Iluvatar CoreX designs general-purpose GPUs (GPGPUs) - the same broad class of parallel-compute chip that Nvidia sells - for training and running artificial-intelligence models. It is a fabless designer: it writes the chip architecture, the instruction set, and the software stack, then has the silicon manufactured by an outside foundry. Its customers are Chinese cloud providers, government-backed AI computing centres, internet companies, and enterprises in finance, healthcare, transportation and education that want to train or serve AI models on hardware that is not made by Nvidia.

The reason a company like this exists in China is almost entirely a geopolitical one. Nvidia dominates AI compute globally, but successive waves of US export controls (2022 onward) have restricted, and by 2025 effectively banned, the sale of Nvidia's most capable AI accelerators - the A100, H100, then the cut-down A800/H800, and finally the H20 - into China. That created a forced, policy-driven demand for a domestic substitute. Iluvatar is one of a small group of well-funded startups racing to fill that gap.

Founding story. The company was founded at the end of 2015 by Li Yunpeng (李云鹏), a returnee who had spent roughly a decade at Oracle in the US as a database R&D director. The founding bet was unusual for its time: rather than build a narrow AI ASIC (the path Cambricon took), Iluvatar committed to a general-purpose GPU - harder to build, but able to run the broad, fast-changing universe of AI and HPC software the way Nvidia's chips do. That decision is the core of the company's identity today.

The company's history is not a clean founder-led arc. Diao Shijing, a former government official, was appointed chairman and CEO in May 2021 and then removed in July 2022 amid an anti-graft investigation - a governance episode worth remembering. The company is today run by Gai Lujiang (盖鲁江), chairman and CEO, who joined in July 2020 and came from a finance and investment background (roughly 17 years across firms including PwC and Deloitte), not a chip-engineering one.

The core value proposition. Iluvatar's pitch is that it was the first Chinese company to mass-produce both a training GPGPU and an inference GPGPU on a 7nm process, and that its chips are CUDA-ecosystem compatible - meaning customers can migrate existing AI workloads with relatively little re-engineering. In a market where the alternative (Huawei's Ascend) runs a different software stack, "you can move your code over" is a real selling point.

What is hard about it. A general-purpose GPU is one of the most complex commercial chips made. It requires a custom instruction set, a SIMT (single-instruction, multiple-thread) compute engine, advanced packaging (Iluvatar's training chip uses 2.5D CoWoS to integrate high-bandwidth memory), and - the part that takes longest - a software stack and compiler that let thousands of real AI models actually run efficiently. Nvidia's true moat is CUDA, fifteen-plus years of software. Replicating even an approximation of that is the central engineering challenge, and the reason all of Iluvatar's R&D spend has, every single year, exceeded its revenue.

A concrete example. A national AI computing centre wants to train and serve large language models without Nvidia hardware. Iluvatar supplies Tianhai (天垓) training accelerator cards, racked into clusters, plus its Zhikai (智铠) inference cards for serving the trained model, plus the Tianshu software stack so the centre's engineers can run mainstream deep-learning frameworks. On a published benchmark, an AquilaCode 7-billion-parameter model trained on a Tianhai-100 cluster reached ~87K tokens/second with a >95% linear scaling efficiency, which Iluvatar positions as broadly comparable to an Nvidia A100 cluster on convergence and speed. That "A100-class, but domestic" framing is the whole business in one sentence.

Iluvatar describes itself as building a "full-stack self-developed" GPU through "open collaboration" - the dual message that it owns its core IP and will be CUDA-compatible enough to plug into existing AI software, rather than asking customers to start over.

Section 2: Business segments

Iluvatar reports as essentially one business - GPGPU compute - but its disclosure and product structure break into three commercially distinct lines. The gross-margin spread between them (below) is wide enough that they behave like separate sub-businesses.

Training products (Tianhai / 天垓 series)

This is the historical core and the highest-margin line. Training GPGPUs are the chips used to build AI models - the most demanding, most performance-sensitive workload, and the one Nvidia's H100/H200 own globally. Iluvatar's Tianhai-100, launched January 2021, was billed as China's first domestically produced 7nm GPGPU. The follow-on Tianhai-150 raises compute density.

• Core capability: a SIMT architecture with a self-defined instruction set, 2.5D CoWoS packaging integrating 32GB of HBM2 memory at ~1.2TB/s bandwidth, and a 7nm process holding ~24 billion transistors. This is the part that took years and is hard to replicate - getting a from-scratch architecture to actually converge large models at high scaling efficiency.
• Competitive position: within China, it competes against Biren's training parts, Huawei Ascend, and (until cut off) Nvidia A100/A800. Its training line carries the richest margins (management cited 53%-61% gross margin), reflecting scarcity of credible domestic training silicon.
• Strategic role: the flagship and the brand. Training leadership is what lets Iluvatar claim "general-purpose" status rather than being boxed as an inference-only vendor.

Inference products (Zhikai / 智铠 series)

Inference chips run already-trained models - a larger, faster-growing, more price-competitive market. The Zhikai-100 (launched May 2022, on Iluvatar's second-generation architecture) delivers ~384 TOPS INT8 / 96 TFLOPS FP16, 32GB HBM2E, 150W, 800GB/s, with broad video-decode support and CUDA compatibility. A smaller Zhikai-50 (16GB, 75W, half-height card) targets lower-power slots.

• Core capability: multi-precision inference (FP32/FP16/INT8), 800+ general instructions, and a claimed 2-3x real-world performance versus some mainstream domestic alternatives, plus heavy media-decode for vision/streaming workloads.
• Why separate: different customer (anyone deploying AI, not just model-builders), different economics. Inference gross margin was the steepest improver - 35.8% rising to 46.7% - as volume and software maturity improved.
• Strategic role: the volume growth engine. Inference demand scales with AI deployment across many industries, so this is where unit shipments multiply.

Solutions, servers and clusters (custom AI compute)

Iluvatar packages its chips into accelerator cards, servers and full clusters, and sells customised AI-compute solutions to data centres and enterprises.

• Core capability: system-level integration plus the Tianshu software stack (X86 and ARM support), which is what turns bare silicon into something a customer's engineers can actually use.
• Why separate: it is a services/integration business with different economics. Its gross margin was the most volatile and the lowest historically - 25.9% rising to 45.7% - because it bundles third-party hardware and bespoke engineering.
• Strategic role: the customer-capture and lock-in layer. Once a buyer runs Iluvatar clusters and software, switching means re-qualifying hardware and re-porting software.

New: edge/endpoint products (Tongyang / 彤央 series)

Launched at the January 2026 roadmap event, the Tongyang line (TY1000 pocket module, TY1100 with a 12-core ARM v9 CPU, TY1100_NX, and TY1200 at ~300 TOPS) extends Iluvatar from the data centre toward edge AI, AI PCs and embodied AI/robotics, spanning a 100-300 TOPS dense-compute range, with TY1000 claimed to beat Nvidia's AGX Orin on several benchmarks. This is a strategic option, not yet a revenue pillar.

Section 3: Products and business detail

Tianhai-100 (training, Jan 2021). China's first domestically produced 7nm GPGPU. 2.5D CoWoS packaging; ~24 billion transistors; 32GB HBM2 at ~1.2TB/s; ~147 TFLOPS FP16 training compute; PCIe Gen4 x16; SIMT scalable compute engine with a self-defined instruction set supporting mixed FP32/FP16/INT8. Demonstrated training of a 7B-parameter model at ~87K tokens/s with >95% linear scaling, positioned as A100-comparable. Cumulative orders for the Tianhai line passed RMB500m early in its life.

Tianhai-150 (training). Density step-up over the 100 (TPP density ~3,040 vs ~2,352), the higher-end training part.

Zhikai-100 (inference, May 2022). Second-generation architecture. ~384 TOPS INT8, ~96 TFLOPS FP16, ~24 TFLOPS FP32; 32GB HBM2E; 800GB/s; 150W; PCIe Gen4 x16; 128-channel multi-format video decode; 800+ general instructions; CUDA-ecosystem compatible. Positioned as the first time a domestic vendor offered a complete "training + inference, cloud + edge" general-compute system.

Zhikai-50 (inference). Half-length/half-height PCIe card, 16GB HBM2E, 75W - for power- and slot-constrained deployments.

Tongyang TY1000 / TY1100 / TY1100_NX / TY1200 (edge, Jan 2026). Modular edge accelerators from pocket-sized (TY1000, 699-pin module) up to TY1200 (~300 TOPS for AI PC and embodied AI), some integrating an ARM v9 12-core CPU. Dense (measured) compute spanning 100-300 TOPS.

Software - the Tianshu stack. A self-developed software stack with CUDA-ecosystem compatibility, supporting mainstream deep-learning frameworks and both X86 and ARM hosts, with integrated debug/optimisation tooling. Management has cited ~95% operator reuse for large models and 1,000+ supported model variants - the metrics that matter for "can a customer actually run their workload."

Manufacturing and the foundry constraint. Iluvatar is fabless. Its 7nm parts were originally fabricated externally (the China GPU cohort historically used TSMC). This is the single most important operational fact about the company: in November 2024 the US tightened controls and TSMC stopped supplying advanced AI chips to Chinese designers, pushing the cohort toward domestic foundries - principally SMIC, whose capability is capped at roughly 7nm. So Iluvatar can keep building 7nm-class parts domestically, but the path to the 5x-12x performance gain needed to reach H200/B200 class - which requires more advanced nodes, better packaging and HBM - runs straight into the export-control wall.

Geographies. The business is overwhelmingly domestic Chinese. Demand is concentrated in national/regional AI computing centres and Chinese cloud, internet, finance, healthcare, transportation and education customers. There is no meaningful export market; the company's reason for existing is import substitution inside China.

Milestones: 2015 founding → 2021 Tianhai-100 (first domestic 7nm training GPGPU) → 2022 Zhikai-100 (inference) → cumulative 52,000+ GPGPU units shipped and 900+ deployments by mid-2025 → Jan 2026 HK IPO → Jan 2026 four-generation roadmap and Tongyang edge launch.

Section 4: Customers

Who buys. Iluvatar served ~290 customers across 900+ deployments by mid-2025 (up from 22 customers in 2022 to 181 in 2024). The base spans national and regional AI computing centres, cloud providers, internet companies, and enterprises in finance, healthcare, transportation, education and autonomous driving. A large share of demand is state-directed - China's "computing-power localisation" push channels government and SOE buyers toward domestic GPUs.

Who decides, and on what criteria. For a national computing centre or large cloud buyer, the decision sits with infrastructure and procurement leadership, often shaped by government localisation mandates. The criteria are: can it train/serve our models at acceptable performance, how painful is the software migration from CUDA, what is the total cost per token, and - increasingly - is the supply chain secure from US action. Iluvatar leans on CUDA compatibility (low migration pain) and its training credentials.

Why they choose Iluvatar. Three concrete reasons: (1) it is one of very few domestic vendors with a credible training GPGPU, not just inference; (2) CUDA-ecosystem compatibility lowers the porting cost relative to Huawei Ascend's separate stack; (3) policy - buyers under localisation directives need a non-Nvidia option, and Iluvatar is an approved, now publicly listed, one.

Switching costs. Real but not absolute. Once a customer ports models to the Tianshu stack and racks Iluvatar clusters, moving away means re-qualifying hardware and re-porting software - meaningful friction. But because Iluvatar deliberately mirrors CUDA, the lock-in is weaker than a fully proprietary stack would create, and a customer could move to another CUDA-compatible domestic vendor more easily than off Nvidia.

Concentration. Historically extreme and now improving. Top-five customers were 94.2% of revenue in 2022, still above 70% in 2024, falling to 38.6% in H1 2025 (the company attributes part of the H1 dip to seasonality, since large buyers concentrate purchases in the second half). The trajectory is healthy, but a business this dependent on a handful of large, often state-linked buyers carries real revenue-timing and customer-loss risk.

Contract structure. Largely project- and order-based hardware sales (chips, cards, servers, clusters) with attached solution/software work, rather than recurring subscription revenue. That makes revenue lumpy and back-half-weighted, and explains the seasonality management flagged.

Section 5: Competitive landscape

Iluvatar sits in the most crowded, most subsidised corner of the Chinese chip industry: the race to be the domestic Nvidia. The structure has three tiers - the incumbent (Nvidia, being squeezed out by policy), the domestic heavyweight (Huawei Ascend), and the venture-funded "four dragons" startups, of which Iluvatar is one.

Where Iluvatar wins: it is one of the few with a mass-produced training GPGPU and a CUDA-compatible stack, and it now has the validation and balance-sheet cushion of a Hong Kong listing. Where it loses: it is small (~3% of domestic-vendor shipments in 2025, ~0.3% of the broader smart-computing market in 2024), still deeply loss-making, and several generations behind Nvidia's frontier - it would need a 5-12x performance jump to reach H200/B200 class, which the foundry constraint makes very hard. Against Huawei it loses on scale, ecosystem heft and supply security; against Cambricon it loses on profitability.

Barriers to entry are genuinely high (architecture, software stack, packaging, foundry relationships, capital - the four dragons collectively burned ~RMB15bn over 2022-2024), which protects the incumbents from new entrants but does nothing to thin the existing, well-funded cohort. This is the key tension: high barriers, yet six-plus serious players chasing a market where one foreign incumbent still held over half the share in 2024. Margins on the solutions line already show the commoditisation pressure.

Market caps are rough peer-size references only, with wide uncertainty given recent listings and volatility; they are not applied to Iluvatar.

The structural shift to watch: domestic vendors collectively pushed Nvidia's China AI-chip share below 60% through 2025, but Nvidia is attempting re-entry with the H200. The competitive question for Iluvatar is therefore two-front - take share from a returning Nvidia and survive a brutally crowded domestic field.

Section 6: Industry

Demand driver. The single dominant driver is China's drive for AI compute self-sufficiency under US export controls. Washington progressively banned Nvidia's A100/H100, then the A800/H800 work-arounds, then the H20, and in November 2024 cut Chinese AI designers off from TSMC's advanced nodes. Every tightening converts directly into demand for domestic GPGPUs. Secondary drivers are the broad AI build-out - large-model training, inference at scale across internet/finance/healthcare, and the emerging edge/embodied-AI wave.

Size and trajectory. China's AI-accelerator / smart-computing market is large and growing fast, but Nvidia still held ~54.4% in 2024 even under restrictions, with the entire domestic-startup cohort under ~3% combined. The domestic share is rising structurally as localisation mandates bite and Nvidia is squeezed. The headline numbers are striking but two-sided: the four dragons generated only ~RMB2.82bn of revenue against ~RMB15bn of losses over 2022-2024 - a market defined more by strategic necessity than by current economics.

Where Iluvatar sits in the supply chain. It is a fabless designer dependent on external foundry (historically TSMC, now domestic SMIC), HBM memory suppliers, and advanced-packaging capacity - all of which sit in the crosshairs of export policy. It owns the architecture and software; it does not own the hardest-to-replace input, leading-edge fabrication.

Import substitution. This is the industry thesis. Nvidia is the incumbent import; domestic GPUs are the substitute; policy is forcing the substitution faster than free-market economics would. Iluvatar is a direct play on that substitution curve.

Regulation. The regulatory environment cuts both ways: US export controls create the demand, but also cap the ceiling by denying advanced nodes, HBM and EDA tools. A future US Entity List designation (peers have faced this) would be the sharpest single regulatory risk.

Cyclicality. Less classic semiconductor cyclicality, more policy- and capex-cycle dependence. Demand tracks government AI-infrastructure spending and localisation directives, and revenue is seasonally back-half-weighted on large-buyer ordering patterns. A tightening of state computing-centre budgets would hit this cohort hard.

Tailwinds: localisation mandates, expanding AI deployment, edge/embodied-AI emergence, Nvidia's policy-driven retreat. Headwinds: foundry and HBM access ceilings, intense domestic competition, Nvidia's attempted H200 re-entry, and the sheer capital intensity of staying in the race.

Section 7: Growth triggers

Caveat: no earnings-call transcripts exist (company listed January 2026). The following are drawn from the post-IPO FY2025 results (~1 Apr 2026), the 26 January 2026 architecture-roadmap investor event, and the IPO prospectus. Forward-looking statements only.

• Four-generation architecture roadmap targeting Nvidia parity and beyond (26 Jan 2026 roadmap event). Tianxuan architecture targeting Blackwell-class in 2026, Tianji exceeding Blackwell in 2026, Tianquan surpassing Rubin in 2027 - products slated to compete with Nvidia H200/B200 across 2026-2028.

  Management framed a cadence in which the 2025 Tianshuo architecture "surpasses Hopper," with successive generations targeting Blackwell parity (2026) and beyond (post-2027 "breakthrough computing chip architecture").

• Tongyang (彤央) edge product line launched (26 Jan 2026). Four new edge/endpoint modules (TY1000-TY1200, 100-300 TOPS) opening AI PC, embedded and embodied-AI/robotics as new end markets beyond the data centre.

• Inference volume ramp (FY2025 results, ~1 Apr 2026). GPGPU shipment volumes rose 168.5% in FY2025, with management attributing FY2025 revenue growth (+91.6% to RMB1.03bn) to wider chip adoption - the volume curve management expects to continue as localisation deepens.

• Customer-base broadening reducing concentration (prospectus / H1 2025). Customer count expanded toward ~290 with 900+ deployments, and top-five concentration fell from 94.2% (2022) toward 38.6% (H1 2025), which management positions as a more durable revenue base going forward.

• IPO proceeds funding next-gen R&D and capacity (IPO, Jan 2026). The ~HK$3.7bn (~US$475m) raise is earmarked to fund the roadmap above - next-generation chip development, software-stack expansion and commercialisation.

• Software-stack maturity as an adoption unlock (26 Jan 2026). Management cited ~95% operator reuse and 1,000+ supported models, framing continued software maturation as the lever that converts pipeline into shipments.

Section 8: Key risks

Foundry and advanced-node access (high probability, high severity). The defining risk. TSMC stopped supplying Chinese AI designers in November 2024; Iluvatar must rely on domestic foundry (SMIC) capped around 7nm, while the roadmap requires far more advanced nodes, HBM and packaging to reach H200/B200 class. The 5-12x performance gap to Nvidia's frontier cannot be closed without inputs policy currently denies. This is a structural ceiling, not a transient supply hiccup.

US Entity List / sanctions escalation (moderate probability, severe). Iluvatar was not identified as currently Entity-Listed in available sources, but several peers have been targeted, and the entire cohort lives under the threat of designation, which would cut off remaining foreign tooling, IP and HBM and could impair foundry relationships. A single regulatory action could materially impair the roadmap.

Persistent heavy losses and cash burn (high probability, moderate-to-severe drag). The company has never been profitable: net losses of RMB554m (2022), 817m (2023), 892m (2024), 609m (H1 2025) and ~RMB1bn (FY2025), with cumulative losses of roughly RMB2.87bn through H1 2025. R&D has exceeded revenue every year. The IPO refills the tank, but the business must keep raising/burning to fund the roadmap, and the path to breakeven is unproven.

Customer concentration and state-demand dependence (moderate probability, moderate severity). Despite improvement to 38.6% top-five in H1 2025, the business still leans on a small set of large, often state-linked buyers, with seasonal back-half ordering. A budget pullback at national computing centres, or the loss of one or two anchor accounts, would show up sharply.

Competition and commoditisation (high probability, moderate severity). Six-plus well-funded domestic players plus Huawei plus a returning Nvidia (H200) chase a market where the domestic cohort is still ~3%. The solutions segment's historically low margins already signal price pressure. CUDA compatibility, while a selling point, also lowers switching costs between domestic vendors.

Governance history (low-moderate probability, reputational). A prior chairman/CEO (Diao Shijing) was removed in July 2022 amid an anti-graft investigation. The current CEO has a finance/investment rather than chip-engineering background. For a company whose entire value rests on out-engineering a brutal field, governance and technical-leadership continuity deserve scrutiny.

Valuation/expectation risk (qualitative). The stock debuted +31.5% and was oversubscribed 414x into a thematic frenzy; expectations are priced for a domestic-Nvidia narrative that the financials (sub-RMB1.1bn revenue, RMB1bn losses, ~3% share) do not yet support. The gap between story and substance is itself a risk.

Section 9: Walk the talk

The five reporting periods used: FY2022, FY2023, FY2024 and H1 2025 (all from the January 2026 IPO prospectus), plus the post-IPO FY2025 annual results (~1 Apr 2026), supplemented by the 26 January 2026 roadmap event. There are no earnings-call transcripts - the company has been public for ~5 months and reports half-yearly - so this section assesses management's track record from disclosed financials and stated commitments rather than from five calls. I flag that limitation plainly; the most recent disclosure (FY2025, ~Apr 2026) is within 90 days of today.

On the numbers that can be tracked across these periods, the operational story has been consistent and delivered. Revenue climbed RMB189m → 289m → 540m (a 68.8% 2022-24 CAGR), H1 2025 grew 64.2% year-on-year to RMB324m, and the first full-year result as a public company landed at RMB1.03bn, up 91.6% - i.e. the growth did not fade after listing, which is the most important early credibility test for a freshly IPO'd name. Unit shipments rising 168.5% in FY2025 corroborates that the revenue is volume-driven, not a one-off large order. Gross margins held in a tight ~49-59% band, and the lower-margin inference and solutions lines improved (inference 35.8%→46.7%, solutions 25.9%→45.7%), which is the right direction and suggests pricing/efficiency claims were not hollow.

Where management's claims are less proven is on the two things that matter most for the long thesis: profitability and frontier performance. The net-loss rate improved markedly (292.3% in 2022 to 165.4% in 2024), which management can fairly point to as progress - but the absolute loss kept growing (to ~RMB1bn in FY2025), so "narrowing losses" is true as a ratio and false as a number. A skeptical reader should hold management to the absolute path. On performance, the January 2026 roadmap promises Blackwell-parity in 2026 and Rubin-beating silicon by 2027 - bold commitments that collide directly with the foundry constraint, and that have no track record yet because they are brand new. These are the promises to watch; there is no history of delivery against them to lean on.

Management's roadmap claim that successive architectures will "surpass Hopper" (2025), reach Blackwell parity (2026) and "surpass Rubin" (2027) is the central testable commitment. With no advanced-node access and a stated 5-12x gap to H200/B200, the gap between this guidance and the manufacturing reality is the single thing to track at each future result.

Assessment: On the growth-and-execution metrics there is a short but clean record of doing what they said - revenue, shipments and margins all moved the promised way, including the first post-IPO result. On the two existential questions (turning a profit, and closing the performance gap to Nvidia under export controls) management is making large promises with no track record to back them, and one of them (frontier parity) sits in direct tension with the supply-chain reality. Read it as: operationally credible so far, strategically unproven - the kind of management that delivers on the controllable and is asking investors to trust it on the uncontrollable.

Section 10: Shareholder friendliness index

Dividends. Iluvatar has paid no dividends and is highly unlikely to for the foreseeable future. It is a freshly listed, pre-profit company that lost roughly RMB1bn in FY2025 and has never earned a profit; there is no DPS history to report and no payout ratio to compute. Capital is being directed entirely into R&D and growth - appropriate for the stage, but the reader should expect zero income.

Buybacks and dilution. The MoatMap disclosure feed records zero buybacks in the trailing ~90 days (since 2026-03-16), and external searches surface no buyback programme - consistent with a company that IPO'd in January 2026 and is burning cash; a loss-making, capital-hungry chip startup repurchasing its own shares months after raising money would be illogical, and none has occurred. On dilution, the direction is firmly the other way: the January 2026 IPO issued ~25.43m new shares, increasing the share count, and the company will likely need further capital to fund its multi-year roadmap, implying ongoing dilution risk rather than share retirement. Across the relevant window the net effect is share creation, not reduction.

Verdict: Hoards Capital (reinvests/burns) - it pays nothing and repurchases nothing, because every dollar is needed to fund an unproven race against Nvidia and a crowded domestic field; rational for the stage, but offers shareholders no capital return and ongoing dilution risk.

Section 11: Insider activities

Per the instruction governing Hong Kong (a gated venue where the HKEX Disclosure-of-Interests portal is access-restricted and web search returns auth-blocked stubs), the MoatMap cross-market disclosure database is the sole source for recent insider transactions here.

MoatMap records 0 insider transactions for 9903.HK over the trailing 12 months (data current 2026-06-13).

This is expected and not a red flag. Iluvatar only began trading on 8 January 2026, so for most of the trailing-12-month window the shares were not publicly listed, and post-IPO insider dealing by directors, executives and pre-IPO shareholders (including cornerstone investors ZTE and 4Paradigm, and venture backers such as Centurium Capital and HongShan) is typically restricted by IPO lock-up undertakings in the first months after listing. There are therefore no open-market directors' or substantial-shareholder dealings to assess for conviction.

Net assessment: No reportable insider buys or sells in the available window - neutral by absence of data, not by evidence of disinterest. There has been no insider open-market buying to read as a bullish signal, and no selling to flag as a concern; the lock-up period largely precludes either. This section should be revisited once lock-ups expire and the first post-IPO HKEX DI filings appear, as early director or cornerstone-investor dealing will then carry real signal.

Section 12: Scenarios

Bull case. US export controls stay tight or tighten further, and Nvidia's H200 re-entry is blocked or throttled - so the policy tailwind that creates Iluvatar's market only strengthens. China's national computing centres keep buying domestic, and Iluvatar's CUDA-compatible stack makes it the lowest-friction migration target, pulling cloud and internet buyers off the fence. The roadmap broadly delivers: the Tianxuan/Tianji generations land close enough to Blackwell-class that Iluvatar graduates from "A100-substitute" to genuinely competitive training silicon, while the Zhikai inference line scales on the deployment wave and the new Tongyang edge line opens AI-PC and robotics as a second growth vector. Customer concentration keeps falling, the lumpy revenue smooths, gross margins drift up as software matures, and absolute losses finally bend toward breakeven as scale arrives. Iluvatar consolidates as one of two or three surviving domestic GPGPU champions in a market it no longer has to share with Nvidia.

Base case. Iluvatar keeps growing revenue fast off a small base as localisation mandates funnel state and SOE demand its way, and shipments keep climbing - but it stays loss-making for years, because the foundry ceiling caps how far the performance can stretch and because six-plus rivals plus Huawei keep pricing competitive. The roadmap slips and compromises against the SMIC-7nm reality: products improve generation-on-generation but remain a notch or two behind whatever Nvidia is allowed to sell, so Iluvatar wins the workloads where "good-enough domestic" matters more than peak performance. It raises capital again to keep funding R&D, diluting holders. It survives and matters, as a structurally subsidised national-champion-tier vendor with respectable revenue and durable losses - a real business propped up by policy rather than a self-sustaining one.

Bear case. Two things go wrong at once. First, the supply chain tightens: a US Entity List designation, or further restrictions on SMIC/HBM, chokes Iluvatar's ability to make competitive parts, freezing the roadmap and stranding the H200/B200 ambitions. Second, the field consolidates against it: better-capitalised rivals (Moore Threads, MetaX, Biren) or Huawei's full-stack scale win the anchor national-centre contracts, and because CUDA compatibility lowers switching costs between domestic vendors, Iluvatar's installed base proves less sticky than hoped. Losses keep widening in absolute terms, capital markets cool on the thematic trade, and a follow-on raise comes at a punitive price or not at all. Add the governance overhang (a prior chairman removed in an anti-graft probe) and a thin technical-leadership bench, and Iluvatar slides from contender to also-ran in a market that can only sustain a few winners.

Sources

• Caixin: Iluvatar CoreX climbs in HK debut, $5.3bn valuation
• Caixin: Iluvatar CoreX seeks $475m in HK listing (cornerstones ZTE, 4Paradigm)
• SCMP: Biren, Iluvatar CoreX post triple-digit growth but losses persist
• MarketScreener: Iluvatar CoreX reports 92% jump in annual revenue
• TrendForce: Iluvatar CoreX 2026-28 GPU roadmap targeting H200, B200
• STCN: Iluvatar four-generation architecture roadmap + Tongyang edge launch
• Wikipedia: Iluvatar CoreX (founding, leadership, products)
• HelloChinaTech: China GPU IPO gold rush - betting billions on sub-3% share
• 美股之家 (mg21): 天数智芯 product/financials profile
• Iluvatar CoreX official product page (Zhikai/Tianhai/Tongyang)
• Asia Times: TSMC 7nm ban targets China's AI chipmakers
• SCMP: China tech firms fear curbs as TSMC stops AI chip supply
• TrendForce: China's GPU trio - Moore Threads, MetaX, Cambricon
• guancha: 国产GPU独角兽天数智芯冲刺港股IPO (margins >50%)

(Section 13 — Further reading — is intentionally omitted: none of SemiAnalysis, Stratechery, or MBI Deep Dives has published an article centrally about Iluvatar CoreX. SemiAnalysis has covered peer Biren and Stratechery covers the China export-control macro, but Iluvatar is not the central subject of any qualifying free piece.)

A brief note on what I could not fully verify: the FY2025 full-year gross margin and cash position were not cleanly disclosed in the sources reachable; the segment gross-margin figures are from H1 2025 prospectus disclosure; and the cumulative loss (~RMB2.87bn) covers 2022 through H1 2025. The MoatMap insider feed and Hong Kong lock-up status mean Section 11 is genuinely empty rather than under-researched. I'd revisit the insider section after the first post-lock-up HKEX DI filings.