Chinese technology giant ByteDance is developing its own central processing units (CPUs) to support its growing AI infrastructure needs. According to Reuters, the move is being driven by surging chip prices and prolonged supply shortages that are constraining the company’s expansion plans.
The move places ByteDance alongside a growing cohort of global hyperscalers, including Alphabet's Google, Amazon, and Microsoft, that have concluded the economics of custom silicon outweigh the complexity of designing it. It also underscores the industry's rapid strategic shift from AI training toward inference, where deployed models perform agentic tasks that place far greater demands on CPUs working in tandem with the Nvidia graphics processors that have dominated the AI boom.
A Market Under Strain
The pivot to inference-heavy workloads has triggered an acute shortage of server CPUs in recent months, turning what was once a commodity component into a strategic bottleneck. The crunch has had knock-on effects throughout the supply chain, elevating the fortunes of incumbent chipmakers Intel and AMD as credible challengers to Nvidia's AI dominance but also driving up costs sharply for the companies most dependent on their products.
ByteDance currently sources CPUs from both Intel and AMD. According to two of the sources cited by Reuters, those vendors have raised prices significantly, with quarter-over-quarter increases ranging from 10 per cent to as much as 35 per cent in recent months, a trajectory that has prompted ByteDance to accelerate its push for in-house alternatives.
Intel warned Chinese customers of server CPU delivery lead times of up to six months earlier this year. The company said last month that demand from AI firms was so strong in the first quarter that it sold even chips it had originally written off. AMD CEO Lisa Su, meanwhile, warned last week that the global CPU market is "tight," with demand outpacing forecasts and supply constraints expected to persist.
ByteDance's In-House Silicon Play
ByteDance, the Beijing-based parent of short video platform TikTok, is targeting deployment of its proprietary CPU in its own servers and data centres to support internal operations, according to the first source cited by Reuters. The push is tied directly to the company's preparations for a massive rollout of agent-based products, including its Coze platform, an AI agent builder it has been expanding aggressively across global markets.
The company has approached several external partners to assist with the effort. Those partners are expected to contribute not only to chip design work but also to help secure manufacturing capacity at foundries, the sources said. The project remains at an early stage.
Two Architectural Bets
According to the sources cited by Reuters, ByteDance is pursuing two chip architecture tracks simultaneously:
- Arm: the instruction set architecture owned by SoftBank, which underpins a growing share of server silicon from Apple, Amazon, and others
- RISC-V: the open-source architecture gaining traction as an alternative to proprietary designs, particularly among Chinese technology companies navigating export controls
Developing two designs in parallel is a common strategy for technology giants, as it allows them to test options before committing to a costly, large-scale manufacturing run. ByteDance is weighing which design best suits its long-term data centre requirements, the sources said. Arm did not respond immediately to a request for comment.
China’s AI Strategy Reset
How Beijing’s industrial AI, patents, and tight governance are shifting global competition beyond consumer chatbots.
The Broader Race for Custom Silicon
ByteDance's move reflects a broader industry reckoning. As inference workloads grow and AI becomes operationalised at scale, the balance of compute demand is shifting in ways that legacy procurement models struggle to accommodate.
Google has been building its own Tensor Processing Units (TPUs) for years; Amazon's Graviton CPU line powers a significant share of its cloud infrastructure; Microsoft has moved to develop its own silicon for Azure. For ByteDance, achieving comparable control over its hardware stack is less about prestige and more about economic and operational necessity, the ability to scale AI deployment without being subject to the pricing power and delivery timelines of third-party vendors.
The company's decision to explore RISC-V in particular will also attract attention, given the ongoing US restrictions on advanced semiconductor exports to China. The open-source architecture theoretically allows Chinese firms to design chips without relying on US-licensed IP, though manufacturing those chips at leading-edge nodes remains constrained by access to equipment from companies including ASML and TSMC.
Redesigning Data Center Cooling
Map AI-driven rack density, building constraints and power limits to direct-to-chip, immersion and hybrid cooling roadmaps.
What This Means for the Industry
The CPU market's sudden tightness, historically the quieter corner of the data centre hardware stack, signals how profoundly inference is reshaping AI infrastructure priorities. For enterprises and cloud providers tracking the AI buildout, ByteDance's entry into custom CPU design is further evidence that vertical integration is no longer optional for hyperscale AI operators. It is quickly becoming the norm.
Comments ( 0 )