Yazılar

ByteDance Shifts Chip Design Staff to Singapore Unit Amid U.S.-China Tensions

/in /tarafından

Chip designers at ByteDance, many based in Beijing and Shanghai, were surprised last week to learn they are officially reporting into a Singapore unit, according to three people familiar with the matter. The change became clear when staff were reassigned into a new group on the company’s internal messaging system.

Analysts suggest the restructuring could help ByteDance navigate U.S.-China trade restrictions on semiconductor access. Since late 2023, U.S. rules have barred mainland Chinese firms from using Taiwan’s TSMC to manufacture advanced AI chips above certain performance thresholds. Shifting oversight to Singapore may allow ByteDance more flexibility in securing partnerships and production.

ByteDance, best known globally for TikTok, has been expanding into proprietary chip design since 2022, developing application-specific integrated circuits (ASICs) to reduce reliance on suppliers like Nvidia. The company has worked with Broadcom on AI processors intended for TSMC fabrication, though it does not currently outsource manufacturing to the Taiwanese firm.

The Singapore entity may be linked to Picoheart, a ByteDance subsidiary registered in December 2023. Picoheart drew notice last year when it acquired a 9.5% stake in Chinese memory chipmaker Innostar. Singapore also hosts TikTok’s CEO Shou Zi Chew and some of ByteDance’s largest data centers.

So far, ByteDance’s chips are limited to inference tasks, such as video decoding and networking, rather than the more computationally intensive AI training workloads where rivals like Alibaba and Baidu have advanced further. Job postings indicate ByteDance is still hiring for its AI chip team as it tries to catch up in the strategic semiconductor race.

OpenAI to Debut First AI Chip in 2026 With Broadcom Partnership

/in /tarafından

OpenAI will launch its first in-house artificial intelligence chip in 2026 through a partnership with U.S. semiconductor leader Broadcom (AVGO.O), according to the Financial Times. The chip will be used internally to power OpenAI’s own AI systems rather than being sold to external customers, people familiar with the matter said.

The move reflects OpenAI’s push to diversify away from Nvidia, whose GPUs currently dominate AI computing, and to lower infrastructure costs amid surging demand for training and running large-scale AI models like ChatGPT. OpenAI has previously collaborated with Broadcom and Taiwan Semiconductor Manufacturing Co. (TSMC) on design and fabrication, while also supplementing with AMD and Nvidia chips.

Reuters earlier reported that OpenAI was finalizing the design of its first custom silicon, to be manufactured at TSMC, with a focus on reducing reliance on outside suppliers. By developing its own chip, OpenAI joins rivals Google, Amazon, and Meta, which have already rolled out proprietary processors to handle escalating AI workloads.

The timing of the news coincides with Broadcom CEO Hock Tan’s announcement on Thursday that the company had secured over $10 billion in AI infrastructure orders from a new unnamed customer, set to drive significant revenue growth in fiscal 2026. Industry watchers say OpenAI could be that customer, given its scale and need for dedicated compute.

If successful, the partnership would not only help OpenAI gain greater control over its AI infrastructure but also cement Broadcom’s position as a leading custom silicon provider in the generative AI era.

Chinese Tech Firms Still Pursuing Nvidia Chips Despite Government Pressure

/in /tarafından

Chinese tech giants including Alibaba (9988.HK) and ByteDance remain eager to secure Nvidia’s (NVDA.O) artificial intelligence chips despite regulators in Beijing discouraging such purchases, according to four sources familiar with procurement talks.

The companies are pressing for assurance that their orders for Nvidia’s H20 model—which regained U.S. approval for sale in China in July—are being processed. They are also closely tracking Nvidia’s development of a more advanced chip, tentatively called the B30A, based on its Blackwell architecture. Sources said the B30A could cost roughly twice as much as the H20’s current $10,000–$12,000 price tag but may deliver up to six times more power, making it an attractive option if cleared by Washington.

Both the H20 and B30A are downgraded versions of Nvidia’s global products, designed to comply with U.S. export restrictions. The issue of whether Chinese firms can access advanced chips remains a central flashpoint in the U.S.–China technology rivalry. While Washington has relaxed some curbs, U.S. President Donald Trump recently struck a deal requiring Nvidia to give 15% of its H20 revenue to the U.S. government.

China, meanwhile, is urging its companies to reduce reliance on U.S. chips. Regulators have summoned Tencent (0700.HK), ByteDance, and others to question their H20 purchases, citing potential information security risks. However, Beijing has not formally banned Nvidia products.

Strong demand persists due to limited domestic chip supply. Products from Huawei and Cambricon (688256.SS) remain constrained and, according to engineers at Chinese firms, perform less effectively than Nvidia’s. Nvidia itself acknowledged rising competition from local rivals but declined further comment.

Uncertainty over its China sales led Nvidia to issue a cautious forecast in August, excluding potential revenue from the world’s second-largest economy. The company’s shares have since fallen about 6%. CEO Jensen Huang has reassured Chinese customers about H20 availability and is reportedly preparing B30A samples for delivery to China as early as September. Nvidia is estimated to hold 600,000–700,000 H20 units in inventory and has asked TSMC to produce more.

Huang has previously said China could represent a $50 billion market for Nvidia if it maintains access to competitive products.