ByteDance Doubles Iluvatar Orders to 100,000 Inference Chips

ByteDance has doubled GPU supply from Shanghai’s Iluvatar CoreX to 100,000 chips in 2026, one person familiar with the allocations said. Most of those parts are meant for inference, the work of answering user queries for Doubao, the company’s consumer chatbot. In June, people close to the talks had put the ByteDance run at least 50,000 chips, and said the terms were still in flux.

If the deliveries land, Iluvatar CoreX becomes ByteDance’s third major Chinese GPU supplier after Huawei and Cambricon Technologies, the Beijing AI-chip designer. ByteDance has also weighed Baidu’s Kunlunxin chips, a line Tencent already uses. ByteDance, Iluvatar CoreX, Baidu, Tencent, Huawei and Cambricon did not comment.

ByteDance Shopped for a Third GPU Maker

The June discussions were about inference silicon, not a full replacement for the most powerful training cards. People familiar with the matter said Iluvatar CoreX, listed in Hong Kong in January 2026 under ticker 9903, had until then sold mainly into government procurement. A hyperscaler order of that size would be a commercial step-change for a firm that reported 1 billion yuan ($148 million) of 2025 revenue, about 90 percent from GPUs.

Huatai Securities, in a note cited alongside those talks, put 2026 revenue at 3.04 billion yuan and said total shipments would jump 139 percent to over 100,000 chips company-wide. The broker put the average selling price of Zhikai inference parts at 12,000 yuan. Iluvatar CoreX shares rose 12 percent in Hong Kong on June 15 after news of the talks spread.

Chinese GPU and AI chipmakers held nearly 41% of China’s AI accelerator server market in 2025, according to industry figures released in April. Nvidia chief executive Jensen Huang has said the firm’s China share has effectively fallen to zero. Tencent chief strategy officer James Mitchell said in May that Chinese AI chips would arrive in large quantities in the second half of 2026.

Chinese AI chips would become available in large quantities in the second half of this year

James Mitchell, Chief Strategy Officer, Tencent, May 2026

Mitchell was talking about supply, not price. Three months later the same domestic names were quoting higher, and Iluvatar was moving more of its book toward one internet customer.

100,000 Chips, Diverted From Internal Stock

On September 10, people familiar with pricing and allocations said Iluvatar CoreX had doubled GPU shipments to ByteDance to 100,000 units in 2026. The company also diverted GPUs earmarked for internal use to help meet that demand, one of them said, as computing constraints intensified.

That 100,000-unit ByteDance allocation is a customer figure. It sits beside Huatai’s separate forecast that Iluvatar’s whole company would ship over 100,000 chips in 2026. If both hold, TikTok’s parent is no longer a trial account. It is the volume line.

THE BYTEDANCE SUPPLY SHIFT

1. January 2026: Iluvatar CoreX lists in Hong Kong after years of mostly state buying.
2. June 15, 2026: People close to the talks say ByteDance is expected to take at least 50,000 chips, mostly for Doubao inference, with terms still open.
3. August 28, 2026: The chipmaker posts first-half revenue of 945.7 million yuan, already close to its full 2025 total.
4. September 10, 2026: A person familiar with allocations says GPU supply to ByteDance has doubled to 100,000 units, with some internal stock redirected.

The June language was “in talks” and “subject to change.” The September language is shipments and a reshuffle of the factory’s own boards. ByteDance still has not described the contract in public, so the 100,000 figure remains a sourced allocation, not a filing.

What Doubao Needs From These Chips

Inference is the cheaper, noisier half of an AI build. Training a large model wants dense interconnects and the strongest cards. Answering millions of chatbot turns wants volume, memory bandwidth, and a price that still works after the query is free to the user. Doubao’s expansion is a query problem, which is why a government-GPU vendor with an inference line was even in the room.

Iluvatar’s own site splits the catalogue that way. The Zhikai series for inference tasks launched in December 2022 as 智铠 100, with mixed FP32, FP16 and INT8 math plus video decode. The Tiangai series for AI training dates to March 2021, when 天垓 100 shipped as the firm’s first cloud training GPU. People describing the ByteDance run said most of the 2026 parts are for inference, not training.

ILUVATAR’S TWO CHIP LINES

• Tiangai (天垓): Training-oriented general-purpose GPUs, first mass product in March 2021, aimed at model development and cluster jobs.
• Zhikai (智铠): Inference-oriented GPUs, first product in December 2022, built for query serving, mixed precision, and lower cost per answer.
• Huatai’s working ASP: 12,000 yuan on Zhikai parts, an estimate from the June note, not a disclosed ByteDance contract price.

That split is why a 50,000-chip, then 100,000-chip, conversation could move without anyone claiming Iluvatar had caught Nvidia on training. ByteDance is buying answers. The prestige race still lives on a different rack.

Huawei, Cambricon, Then Iluvatar

People who track ByteDance’s domestic board mix still rank Huawei first, Cambricon second, and Iluvatar CoreX third. That order did not flip when the Iluvatar number went to 100,000. It did pull a former state vendor into the same conversation as two much better-known names, and it left a fourth door open at Baidu.

Kunlunxin is Baidu’s chip unit. Tencent is already a customer, one of the June sources said, which is why ByteDance looking at the same parts is a peer copy as much as a technical bet. A four-vendor stack is also a bargaining tool. No single Chinese designer yet offers Nvidia’s software glue, so the buyer spreads risk across Huawei’s Ascend line, Cambricon’s 590-class parts, Iluvatar’s Zhikai boards, and, if talks go somewhere, Kunlunxin.

BYTEDANCE’S CHINESE GPU BENCH

ByteDance is not picking a national champion. It is stacking whoever can ship inference boards this year, which is a colder test than a launch-day FLOPS slide. The same pattern showed up in the crowd around the June leak: the company was described as talking to almost every maker on the market, domestic and otherwise, because it has to serve Doubao under Chinese rules and still train large models where the best foreign cards can be legally used.

A January Listing Meets a Volume Customer

Iluvatar CoreX went public in Hong Kong in January 2026. Eight months later its interim results to June 30, 2026 showed how fast a government-GPU shop can re-rate once internet firms start buying. Revenue was 945.7 million yuan, up 191.6 percent from 324.3 million yuan a year earlier. Sales of GPGPU products were 916.1 million yuan, up from 276.8 million yuan.

The profit line is messier than the top line. Gross profit was 162.6 million yuan, against 162.4 million yuan a year earlier, a 0.1 percent rise the company itself recorded. Net profit was 106.3 million yuan, after a 609.3 million yuan loss in the first half of 2025. Adjusted net profit, a non-HKFRS figure the company publishes, was 246.0 million yuan. GPGPU product gross profit was 401.8 million yuan, up 188.5 percent, which means the overall gross-profit stall sits in mix, costs, and other lines, not in the core board business going backwards.

ILUVATAR COREX, FIRST HALF

• Revenue: 945.7 million yuan in H1 2026, versus 324.3 million yuan in H1 2025, up 191.6 percent.
• GPGPU sales: 916.1 million yuan, versus 276.8 million yuan a year earlier.
• Gross profit: 162.6 million yuan, barely changed from 162.4 million yuan, as the company reported a 0.1 percent rise.
• Net profit: 106.3 million yuan, after a 609.3 million yuan loss a year earlier; adjusted net profit 246.0 million yuan.

Huatai’s 3.04 billion yuan full-year revenue sketch now has a first-half base that already matches 2025. The ByteDance allocation is the demand story underneath that math. It is also concentration risk: a single internet buyer large enough to take internal boards off the company’s own floor can cut as fast as it ramps.

Grey-Market Memory Feeds the Price Jump

The same September conversations that raised ByteDance’s Iluvatar draw also said Chinese AI chipmakers had lifted prices. Huawei, Cambricon, MetaX and Iluvatar CoreX all moved, three people familiar with the pricing said, because high-bandwidth memory has become the scarce part inside every card.

HBM is stacked memory that lets a GPU pull data fast enough to keep the math units busy. SK Hynix, Samsung and Micron dominate the advanced market. Washington’s December 2024 high-bandwidth memory controls added HBM stacks to ECCN 3A090.c, aimed at slowing Chinese attempts to build advanced AI chips with imported memory. People tracking Chinese production said designers now lean on grey-market HBM that costs several times what buyers outside China pay, and that those bills flow straight into finished cards.

Huawei lifted the indicated price of its Ascend 950DT accelerator card, a board that pairs a processor with memory, to above 250,000 yuan. That is 20 percent to 50 percent above quotes given to customers two months earlier, depending on contract terms, two people said. Huawei has said the 950DT, its most advanced AI chip, will be available in the fourth quarter of 2026. Cambricon repriced its planned 690 chip 20 percent to 30 percent above levels indicated two months earlier, without giving a yuan figure. MetaX and Iluvatar CoreX made similar increases, the people said.

INDICATED CARD PRICES IN SEPTEMBER 2026

Those increases land on a China AI chip market the same people put at $50 billion, a void Nvidia’s restricted parts left and that domestic vendors have been filling. The fill is real. It is also more expensive than the substitution slides implied, because memory, not just logic, is on a control list. Huawei has said the 950 series will use two in-house HBM designs, HiBL 1.0 on the 950PR and HiZQ 2.0 on the 950DT, without saying where that memory is made.

Training Still Uses the Chips China Cannot Buy

None of this makes Zhikai a stand-in for a Blackwell-class training GPU. Inference is where Chinese designers have the cleaner shot, because the job is narrower and the volumes are large enough to pay for a second or third source. Training still wants the parts Washington has made hardest to buy, which is why ByteDance can take 100,000 domestic GPUs for Doubao answers and still need foreign silicon for the heavy model work.

The 100,000-chip run is local silicon bought against a global GPU queue. That is capex substitution at Doubao scale, and it is also the first time Iluvatar’s inference line has been described as a hyperscaler production input rather than a government catalogue item. Huawei remains the largest Chinese supplier to ByteDance. Cambricon is still second. Iluvatar is the name that moved, and the name that had to raid its own inventory to keep up.

Huawei’s 950DT is due in the fourth quarter of 2026 at the new indicated price. Iluvatar’s 100,000-unit ByteDance allocation is a 2026 figure, not a filed contract. Grey-market HBM still sets the floor under both. ByteDance has not said how many of those boards are already in racks, or whether Kunlunxin joins the list.