Beijing, June 15, 2026: The deal has not been signed. The contracts are not final. And yet the moment Reuters published its exclusive on the morning of June 15, 2026, the ByteDance AI chips story stopped being speculation and became market moving intelligence. Shares of Shanghai based Iluvatar CoreX surged as much as 12% intraday on the Hong Kong Stock Exchange before settling back by midday.
- FROM GOVERNMENT CONTRACTS TO HYPERSCALER: WHO IS ILUVATAR COREX?
- THE ZHIKAI CHIP, THE 50,000 UNIT ORDER, AND THE DOUBAO EQUATION
- BYTEDANCE AI CHIPS: THE THREE TRACK STRATEGY THAT IS REWRITING CHINA’S SEMICONDUCTOR PLAYBOOK
- ILUVATAR COREX’S REVENUE FORECAST AND THE MARKET THAT IS EMERGING
- THE POLICY ENGINE BEHIND THE COMMERCIAL DEAL
- THE BOTTOM LINE: WHAT THIS DEAL ACTUALLY MEANS, AND WHAT MOST COVERAGE WILL MISS
- FREQUENTLY ASKED QUESTIONS
The money moved before the ink dried because investors who understand the structural reality of China’s AI semiconductor market knew what the headline actually meant: ByteDance the TikTok parent and one of the most capital intensive AI companies on the planet has decided that domestic chips are no longer the contingency plan. They are the plan.
According to two sources familiar with the matter cited exclusively by Reuters, ByteDance is in active talks with Shanghai based Iluvatar CoreX to purchase AI chips for inference workloads. The same sources confirmed that ByteDance is simultaneously considering a parallel arrangement with Baidu’s chip subsidiary, Kunlunxin. If the Iluvatar CoreX deal is finalized, the company would become ByteDance’s third major domestic GPU supplier, joining Huawei and Cambricon.
Neither ByteDance, Iluvatar CoreX, Baidu, nor Tencent responded to Reuters’ requests for comment. Both sources declined to be named as the talks are not public. This is the story behind the story.
FROM GOVERNMENT CONTRACTS TO HYPERSCALER: WHO IS ILUVATAR COREX?
To understand why this deal matters, you first need to understand what Iluvatar CoreX actually is and what it has been until now. Founded in 2015 and headquartered in Shanghai, Iluvatar CoreX is China’s first company to achieve mass production of a general purpose GPU platform capable of supporting both AI training and inference workloads, according to TrendForce’s January 2026 reporting.
Its product portfolio runs two distinct lines: the Tiangai series, engineered for AI training, and the Zhikai series, designed specifically for inference tasks the workload at the center of this ByteDance discussion, per the company’s own website.
The company’s flagship TianGai 100 operates on 7nm class process technology and is positioned to compete with Nvidia’s A100 and A800 processors in AI training and inference applications, according to CoinDesk’s reporting of the company’s technical specifications. In January 2026, the company disclosed plans to unveil a three generation product roadmap covering 2026 through 2028, with products reportedly targeting Nvidia’s H200 and B200 the current frontier of GPU performance according to a TrendForce report citing ijiwei.
The financial trajectory is striking. According to the company’s own prospectus data reported by uSMART at the time of its Hong Kong IPO, Iluvatar CoreX recorded revenues of approximately 189 million yuan in 2022, 289 million yuan in 2023, and 540 million yuan in 2024. Then came 2025: the company reported 1 billion yuan ($148 million) in full year revenue, with approximately 90% derived from GPU sales, per Reuters. That is a near doubling of the prior year’s revenue, and it came almost entirely from a single product category.
The company reached unicorn status a valuation exceeding $1 billion in 2024, per Wikipedia’s sourced entry on the company. That milestone arrived before the Hong Kong listing, before the ByteDance talks, and before the broader investor recognition now being priced into the stock.
Iluvatar CoreX listed on the main board of the Hong Kong Stock Exchange on January 8, 2026, with an IPO price set at HK$144.60 per share, according to uSMART’s coverage of the listing. The IPO sponsor was Huatai Financial Holdings (Hong Kong) Limited the same institution that would later produce the bullish research note projecting 3.04 billion yuan in 2026 revenue.
The company priced into a receptive market: per MEXC’s reporting of the January 8 trading session, Chinese AI and chip firms including Iluvatar CoreX closed above their Hong Kong IPO prices, with the cohort collectively raising HK$9.3 billion ($1.19 billion).
According to CBInsights’ sourced reporting, Iluvatar CoreX became the second homegrown GPU company to go public in Hong Kong after Biren, arriving as part of a wave of Chinese chip and AI firms accessing public capital markets in early 2026.
Here is what makes the ByteDance development structurally significant beyond the revenue headline: until now, Iluvatar CoreX has mainly supplied government procurement projects, according to one of the Reuters sources. A ByteDance order of at least 50,000 chips the volume cited by Reuters’ sources would represent the company’s first major commercial scale deployment inside China’s consumer AI infrastructure. That is not an incremental contract. It is a category change.
THE ZHIKAI CHIP, THE 50,000 UNIT ORDER, AND THE DOUBAO EQUATION
The Reuters sources specified that the potential Iluvatar CoreX chips are intended for AI inference workloads a distinction that carries enormous commercial weight and is not sufficiently explained in most coverage of this story.
AI inference is fundamentally different from AI training. Training is the process of building a model: it demands the most powerful, most expensive, most export restricted chips in existence Nvidia’s A100, H100, and H200 class hardware. Inference is what happens after training: it is the real time process of answering queries, generating responses, processing images, and delivering outputs to end users at scale. Inference does not necessarily require frontier class silicon. It rewards efficiency, volume, cost per query, and supply chain reliability.
This is the context in which Iluvatar CoreX’s Zhikai series becomes a logical commercial fit for ByteDance. The Zhikai line is designed specifically for inference, according to the company’s website. The Huatai Securities research note cited by Reuters estimated the Zhikai inference chips carry an average selling price of 12,000 yuan, or approximately $1,775 each. At 50,000 units, that implies a potential order value in the range of 600 million yuan roughly 60% of Iluvatar CoreX’s entire 2025 annual revenue from a single customer relationship.
The end application driving this demand is Doubao, ByteDance’s flagship AI chatbot. Doubao was launched in August 2023 and, per Caixin Global’s May 2026 reporting, reached 345 million monthly active users by March 2026, having added 100 million new users in Q1 2026 alone. According to 36Kr’s December 2025 exclusive, Doubao’s daily active users exceeded 100 million the fastest any ByteDance product has achieved that threshold in the company’s history, and with the lowest user growth and marketing costs of any ByteDance product that has crossed 100 million DAU.
The token consumption required to support that user base is staggering. According to data published by China Daily in May 2026 and citing the National Data Administration, daily token usage across Doubao’s platform soared from just over 1 trillion tokens at the start of 2025 to 100 trillion tokens by the end of the year. By early 2026, reporting from Hello China Tech placed the daily token consumption figure at 120 trillion described as a thousandfold increase since Doubao’s commercial launch in May 2024.
China’s broader AI chatbot market context amplifies the demand picture further. Per AI Wiki’s sourced market data, China’s AI chatbot market served more than 700 million monthly active users across all platforms in early 2026, more than double the 2024 figure reflecting both the maturation of domestic models and the rapid uptake of free to use AI assistants by mainstream consumers. Doubao is the category leader inside that 700 million user pool, and its lead is widening.
Every one of those 120 trillion daily tokens requires inference compute. That is the demand signal driving the ByteDance–Iluvatar CoreX discussions. It is not abstract technology policy. It is the arithmetic of running China’s most used AI chatbot at a scale that requires a dedicated, domestically sourced, reliable inference hardware supply chain.
BYTEDANCE AI CHIPS: THE THREE TRACK STRATEGY THAT IS REWRITING CHINA’S SEMICONDUCTOR PLAYBOOK
ByteDance’s engagement with Iluvatar CoreX does not exist in isolation. It is one visible element of a three track hardware strategy that the company has been constructing systematically since US export controls began tightening on advanced semiconductor sales to China.
Track One: Domestic GPU Procurement at Scale. If the Iluvatar CoreX deal closes, it joins Huawei and Cambricon as ByteDance’s confirmed domestic GPU suppliers, per Reuters’ sources. ByteDance is simultaneously evaluating Baidu’s Kunlunxin chips a product line that Reuters’ sources confirmed Tencent is already using as a customer.
The parallel evaluation of multiple domestic suppliers is not indecision. It is supply chain architecture: distributing dependence across the domestic GPU ecosystem so that no single vendor’s capacity constraint or product roadmap issue can interrupt ByteDance’s AI infrastructure buildout.
Track Two: In House Silicon Development. Per a Reuters exclusive from February 11, 2026, ByteDance is developing its own proprietary AI chip and was in talks with Samsung Electronics to manufacture it. According to two sources cited in that Reuters report, ByteDance aimed to receive sample chips by the end of March 2026, with plans to produce at least 100,000 units of the chip designed for AI inference tasks in 2026.
One of the Reuters sources stated that ByteDance is targeting a progressive scale up to as many as 350,000 units. This is vertical integration: owning the silicon stack so that geopolitical export restrictions cannot constrain your production roadmap.
Track Three: Foreign Silicon Where Accessible. ByteDance has also budgeted aggressively for foreign hardware wherever regulatory pathways permit. According to reporting aggregated by financial platforms, the company has preliminarily budgeted approximately 160 billion yuan ($23 billion) in 2026 capital expenditure, with a significant portion allocated to semiconductor procurement, per AI Wiki.
The company has separately committed approximately $7 billion to renting Nvidia GPUs through cloud services outside China as a regulatory workaround, per the same source. It has also planned approximately $5.7 billion in Huawei Ascend processor orders for 2026, according to reporting by MEXC. This is not a company hedging its bets. This is a company building parallel highways because it cannot afford to have any single road closed.
ILUVATAR COREX’S REVENUE FORECAST AND THE MARKET THAT IS EMERGING
The financial implications for Iluvatar CoreX, if the ByteDance deal is finalized, extend well beyond the immediate order value. Huatai Securities, the company’s own IPO sponsor, projects Iluvatar CoreX’s 2026 revenue at 3.04 billion yuan ($449.8 million), per Reuters’ reporting of the research note. That figure would represent a 204% increase over the 1 billion yuan reported for 2025. The same Huatai note projects total chip shipments will jump 139% to over 100,000 units in 2026.
Consider what those projections mean in context. As of June 30, 2025 data from Iluvatar CoreX’s own prospectus, as reported by uSMART the company had delivered more than 52,000 units to over 290 customers and completed more than 900 deployments across sectors including finance, healthcare, and transportation. A ByteDance order of 50,000 units alone would nearly equal the company’s entire prior cumulative delivery volume and that is before any additional commercial customers.
The market is watching this trajectory with urgency, not patience. When Reuters published its exclusive on June 15, the 12% intraday surge in Iluvatar CoreX’s Hong Kong shares reported by EconoTimes was the market quantifying the significance of the transition: from state adjacent GPU vendor to commercial AI infrastructure supplier at hyperscaler scale.
That transition, if sustained, has sector wide implications. Baidu has already announced plans to spin off Kunlunxin for an independent listing, per CBInsights’ reporting. Alibaba is reportedly considering a similar spin off for its chip design arm, T Head, according to the same source. The Chinese domestic GPU sector is not just producing chips it is restructuring itself as a capital markets investment category, with the explicit expectation that commercial demand from ByteDance, Tencent, and their peers will underwrite the valuations.
THE POLICY ENGINE BEHIND THE COMMERCIAL DEAL
No serious analysis of this transaction is complete without acknowledging the structural policy environment in which it is occurring. US restrictions on advanced chip exports to China have tightened progressively since 2022, with Nvidia’s most powerful processors repeatedly caught in the regulatory crossfire, as CoinDesk reported in its coverage of the ByteDance Iluvatar discussions.
The Chinese government has responded by actively advocating for semiconductor self reliance, directing resources into domestic chip development and encouraging China’s largest technology firms to localize their supply chains, per the same source.
The commercial outcomes now visibly following from that policy posture are exactly what Beijing intended: Tencent using Kunlunxin chips. ByteDance evaluating Iluvatar CoreX and Kunlunxin simultaneously. Huawei Ascend processors ordered at multi billion dollar volumes. The policy incentive and the commercial logic have converged.
Doubao’s paid subscription announcement three tiers at 68 yuan, 200 yuan, and 500 yuan per month, per Caixin Global’s May 2026 reporting adds a revenue dimension to the compute equation. ByteDance stated that some free services will continue while paid tiers target complex productivity use cases.
But the commercialization of Doubao changes the inference economics: if ByteDance is generating per user subscription revenue from Doubao at scale, the cost of inference hardware becomes a directly attributed cost of goods sold, not an R&D line item. That changes how aggressively the company will invest in inference chip procurement and how price sensitive it will be to domestic versus foreign silicon pricing.
At 12,000 yuan per Zhikai chip versus the cost of equivalent Nvidia hardware at export compliant specifications, the domestic pricing advantage is substantial. The supply chain reliability advantage is non negotiable.
THE BOTTOM LINE: WHAT THIS DEAL ACTUALLY MEANS, AND WHAT MOST COVERAGE WILL MISS
Most reporting on the ByteDance–Iluvatar CoreX talks will frame this as a supply chain story. It is not primarily a supply chain story. It is a market legitimization event for the entire Chinese domestic GPU sector, and the difference matters enormously for investors assessing what comes next. Here is what most coverage will miss.
Iluvatar CoreX has existed for over a decade. It has shipped hardware to 290 customers. It has listed publicly in Hong Kong. It has a detailed product roadmap targeting Nvidia’s B200 by 2028. None of those facts, individually or together, changed the market’s assessment of the company’s commercial ceiling because all of those customers were government procurement projects, and government contracts do not prove commercial scalability.
ByteDance is a commercial customer. It is not buying chips to satisfy a state mandate. It is buying chips to run Doubao at 120 trillion tokens per day and growing, for 345 million monthly active users who are now being asked to pay subscription fees. That is a P&L driven procurement decision, and a P&L driven customer validates a vendor’s commercial viability in a way that no government contract ever can.
When the second largest social media company by revenue ByteDance reportedly generated approximately $155 billion in 2024 revenue, per Bloomberg figures cited by Sacra decides your chip is good enough to run its flagship AI product, the entire buyer community takes notice. The Tencents, the Alibabas, the Baidu cloud units, the second tier AI application companies scaling up their inference workloads: they are all watching this transaction to calibrate whether domestic GPU silicon has crossed the commercial readiness threshold.
The Reuters sources were explicit that the deal details are not final and subject to change. That caveat is real and investors should hold it. But the strategic signal has already been sent. The 12% share surge happened before any contract was signed, because the market understands that in the China AI semiconductor sector, the moment a company of ByteDance’s scale and sophistication enters due diligence with a domestic GPU maker, the validation has already begun.
China’s domestic GPU sector is no longer a policy project waiting for commercial customers. It is a commercial market, and it is opening now. Investors who understand that distinction and position accordingly will not be waiting for the press release.
FREQUENTLY ASKED QUESTIONS
FAQ 1: Why is ByteDance buying chips from Iluvatar CoreX instead of just using Nvidia?
The US initially restricted Nvidia from selling its most advanced AI chips to China to prevent the Chinese military and surveillance agencies from accessing cutting edge processing power. These restrictions were first imposed in 2022 under President Biden and were expanded by President Trump in 2025. While some restrictions have since been eased as of May 2026, Nvidia can sell its H200 chip to China, and AMD can sell its MI308 the regulatory environment remains volatile.
Chinese technology companies reportedly placed orders for more than 2 million H200 chips for 2026, but those shipments come with strict conditions: third party testing in the US before export, a volume cap limiting China bound shipments to 50% of domestic US sales, and a 25% tariff attached to each shipment. Sourcing domestically eliminates all of that uncertainty entirely.
FAQ 2: What exactly is an “inference chip,” and why does ByteDance need 50,000 of them?
Inference workloads involve answering queries and are different from AI model training, which tends to use the most powerful chips. In simple terms: training builds the AI model, inference is the model doing its job in real time every time a user types a message into Doubao and gets a response, that is an inference computation. Doubao’s daily token consumption now runs at roughly 120 trillion, a thousandfold increase since its May 2024 commercial launch. At that scale, inference chip procurement is not optional. It is the direct cost of keeping the product running.
FAQ 3: Is Iluvatar CoreX’s chip actually good enough to replace Nvidia hardware?
Iluvatar CoreX is China’s first leading company to achieve mass production of a general purpose GPU platform supporting both AI training and inference. Within the TianGai line, the TianGai 100 delivers a TPP performance density of 2,352, while the TianGai 150 reaches 3,040, still trailing Nvidia’s A100 by a wide margin.
However, Iluvatar CoreX is expected to deliver a substantial performance uplift with its next generation GPUs, targeting Nvidia’s H200 and B200 for 2026–2028 release. Such gains are theoretically achievable through advances in architecture, process technology, and multi chip packaging. For inference specifically not frontier training the performance gap matters considerably less than cost, volume, and supply reliability.
FAQ 4: What does this deal mean for Iluvatar CoreX’s stock and revenue outlook?
Iluvatar CoreX shares rose 12% in Hong Kong following the Reuters report. On the fundamentals side, Iluvatar CoreX’s revenue is projected to reach 3.04 billion yuan ($449.8 million) this year, with total shipments expected to jump 139% to over 100,000 chips, according to a Huatai Securities research note. The broker estimated the Zhikai inference chips had an average selling price of 12,000 yuan, or about $1,775 each.
Prior to this news, as of June 12, 2026 three days before the Reuters exclusive Iluvatar CoreX’s market capitalization stood at $17.2 billion, per PitchBook data. A confirmed ByteDance order at 50,000 units would represent approximately 600 million yuan in revenue from a single customer equivalent to roughly 60% of the company’s entire 2025 annual revenue.
FAQ 5: Is the ByteDance Iluvatar CoreX deal confirmed?
The details of the potential deals are not final and still subject to change, the sources said. ByteDance, Iluvatar CoreX, Baidu and Tencent did not respond to requests for comment. The information was first reported by Reuters on June 15, 2026, citing two sources who declined to be named as the talks are not public. No official announcement has been made by any of the companies involved. Investors should treat this as a credibly sourced indication of strategic intent, not a confirmed transaction.
Connect With Us On Social Media [ Facebook | Instagram | Twitter | LinkedIn ] To Get Real-Time Updates On The Market. Entrepreneurs’ Diaries Is Now Available On Telegram. Join Our Telegram Channel To Get Instant Updates.
Isabella is a global business journalist and former McKinsey analyst from Brazil. She brings sharp insights on economic shifts, policies, and founder journeys from around the world.
Luca is a tech ethicist from Italy exploring disruptive innovation through a human lens—from AI to biotechnologies to decentralization.
