On May 28, 2026, two seemingly unrelated news stories were pushed into the spotlight at the same time: Reuters and several media outlets cited unnamed sources saying that ByteDance is developing its own or customized CPUs, planning to deploy them in its own servers and data centers to hedge against soaring chip prices and long-term supply shortages that are hampering its expansion of AI infrastructure; on the same day, the stock of Zhizhu (02513.HK) surged more than 14% during intraday trading, with Bitget reporting a price of around $206.6, setting a new historical high, as capital responded with substantial buying to the rising sentiment surrounding computing power assets. Research briefs indicated that ByteDance's self-developed CPU plan is still in its early stages and has not yet been officially confirmed by the company, but it has already been interpreted as a signal: as the industry shifts from a "training-centric" to an "inference-centric" focus, the long-standing tension in GPUs is now spilling over to CPU shortages that have started to emerge in recent months, sharply raising the importance of general computing power. Google's self-developed TPU and Amazon's Graviton have already proven that large tech companies have both the motivation and ability to vertically integrate chip design, and ByteDance's follow-up actions, alongside Zhizhu's stock price surge, point to the same conclusion — in the era of AI inference, the pursuit of self-developed CPUs and capital around computing power targets is jointly rewriting the rules of the game regarding who controls computing power and at what price it is allocated.
Chip Crisis: Expensive Computing Power Forces Big Companies to Develop in-house
When GPUs were pushed to the forefront during the training phase, with prices and capacities tightening at high levels for years, computing power for internet companies was no longer a "general resource" that could be replenished at any time but became a primary constraint written into business plans. With the large-scale rollout of big models and a surge in inference requests, the industry focus shifted from "training the model" to "how to keep it online for global users," raising the importance of CPUs during the inference phase swiftly. Research briefs mentioned that CPU shortages have begun to appear in recent months, with computing power tension spilling over from high-end GPUs to general-purpose CPUs, which were once seen as basic components — prices, delivery cycles, and allocation orders have all become strategic variables for companies.
In this supply-demand environment, ByteDance's situation has been called out by reports from Reuters and others: soaring chip prices and long-term supply shortages are practically limiting its expansion plans. For ByteDance, which heavily relies on computing resources to drive recommendation algorithms, content distribution, and AI businesses, this means that every new GPU and every additional CPU cabinet must be weighed repeatedly between costs and supply scheduling, forcing the pace of AI infrastructure construction to yield to upstream capacity. Therefore, when the same batch of reports linked ByteDance's development of its own or customized CPUs directly to its growing demand for AI infrastructure, self-developed or deeply customized hardware was no longer just a side project driven by engineers' interests, but a core strategic option to reclaim control over computing power and hedge against long-term shortages.
ByteDance Builds Chips: Moving CPUs into Its Data Centers
On May 28, Reuters and several media outlets almost simultaneously threw out similar statements: multiple unnamed insiders indicated that ByteDance is advancing a plan for self-developed or deeply customized CPUs, a project that is still in its early stages, with no product details or official roadmap, only a shared consensus that "it is in progress." The research brief presented a more restrained perspective — this is an exploratory self-developed CPU project, with all current information coming from insider reports rather than company announcements, implying that it has moved beyond the planning stage but has yet to enter the stage of external proclamation.
Despite its low profile, the planning direction is not ambiguous. The brief mentioned that ByteDance's goal is to integrate this CPU directly into its own servers and data center cabinets to support internal operations and an expanding AI infrastructure, rather than creating a "commodity chip" for external sales. For a company that heavily relies on recommendation algorithms in short videos and information feeds and processes vast amounts of data daily, this is more like building a "self-sufficient power plant" for its computing power: on one hand, the soaring chip prices and long-term supply shortages have been pointed out in multiple reports as key constraints on its expansion pace. Betting on self-developed CPUs can lock in part of computing power costs on a controllable track over the coming years; on the other hand, when general-purpose CPUs also show signs of shortages, continuing to rely entirely on external supplies equates to handing over business control to upstream capacity schedules. Proactively mastering key infrastructure has become a necessary defense for ByteDance in the era of AI inference.
The Rise of the Inference Era: CPUs Become the New Computing Power Bottleneck
In the previous cycle dominated by "large model training," computing power was almost synonymous with GPUs. Deep learning training requires repeated large-scale parallel floating-point computations on massive datasets, making GPUs naturally suited for this highly parallel matrix computation. Thus, they became the absolute powerhouses during the training stage; this has been an unspoken consensus in the industry. However, when models transition from the lab to online usage, the computing power form in large-scale inference scenarios begins to reverse: when faced with massive user requests, the system not only has to run the model itself but also handle complex preprocessing and business logic, which have long been carried out by CPUs within cloud computing architectures; they are responsible for receiving requests, scheduling tasks, and assembling upstream and downstream services, turning into an indispensable "traffic hub" in the entire inference link.
Research briefs cite related reports noting that ByteDance's self-developed CPU rhythm is interpreted in a larger industry context — the industry is accelerating its transition to the "inference" phase, during which demand for CPUs has significantly increased, and CPU shortages have even emerged in recent months. In other words, computing power tensions are no longer just a matter of GPU prices and supplies but are beginning to spill over to general-purpose CPUs, with new bottlenecks appearing within inference infrastructure itself. In this structural shift, whoever holds enough scale and sustainable, controllable CPUs and general computing power will master pricing and pacing power in the next phase of the AI infrastructure competition.
Zhizhu's Stock Price Hits New High: Capital Pursues Computing Power Stories
On May 28, the logic of computing power rapidly materialized in the capital market into a single stock. Zhizhu (02513.HK) saw its stock price surge over 14% during intraday trading, with Bitget reporting a price near $206.6, hitting a historical high; the stock's momentum almost continually rose, becoming the absolute focus of the AI sector on that day. The market sentiment was very clear: under the narrative of "who controls general computing power," any target linked to upstream computing resources would be repeatedly confirmed by capital on days filled with news.
The coincidence of timing further reinforced this imagination. On the same day that news of ByteDance’s self-developed or customized CPU plan was disclosed by Reuters and other media, the research brief pointed out that the unusual movements in Zhizhu's stock price coincided with this news appearing within the same time window, possibly reflecting the market's optimistic sentiment towards upstream computing power segments in the AI industry chain. However, the brief itself did not provide direct catalysts such as being included in indices, signing large orders, or landing specific projects but only confirmed the correlation between synchronized stock price strengthening and rising computing power sentiment. In the absence of more granular information, attributing a high volatility trend simply to a single piece of news carries the risk of treating emotion as a certitude. The true factors driving this recent surge still await more information to separate noise from signal.
Big Companies Compete in Self-Development: The Next Round of Computing Power Competition Has Begun
When placing ByteDance's recent exposure regarding self-developed or customized CPUs back on a longer timeline, it is not an isolated event but part of catching up with a path already traversed by overseas giants. Google bet on self-developed TPUs many years ago, locking core training and inference tasks within its own data center system; Amazon reshaped the cost-performance of cloud general computing power with its Graviton series; Apple fully transitioned its Macs and mobile devices to in-house developed M-series and A-series chips, proving that as long as the scale is adequate and the software ecosystem is solid, achieving chip design is not an "insurmountable barrier." Following these examples, ByteDance's move towards self-developed or deeply customized CPUs today essentially represents a delayed but inevitable follow-up in the competition for computing power dominance in the era of AI.
Trends indicate that moving forward, whether cloud vendors or AI companies centered around models, it will be difficult to fully entrust their fate to a single upstream supplier. On the CPU side, under the pressure of expanded inference demand, self-development or joint customization with design firms will become a line to hedge against the shortage of general chips; on the accelerator card side, there will be more narrow and deeper optimizations around specific models and business scenarios to lower long-term unit computing costs. The research brief gives the judgment that ByteDance's CPU project is still in its early stages and is years away from mass production and scale deployment due to extended R&D and validation cycles, and semiconductor manufacturing, from design to mass production, involves complex processes such as outsourcing and packaging, which means that even more self-development projects in the short term cannot immediately resolve soaring prices and supply tensions. They may only gain a small amount of negotiating leverage at the table, but once these projects truly land in the medium to long term, whoever holds the sustainable supply of proprietary computing power will have a stronger position in rewriting the rules of the next price competition.
Join our community to discuss and grow stronger together!
On-chain Telegram Community: https://t.me/AiCoinWhaleData
On-chain Community: https://www.aicoin.com/link/chat?cid=N6OVMor5g
AiCoin On-chain Twitter: https://x.com/aicoinwhaledata
AiCoin Exclusive Hyperliquid Benefits: https://app.hyperliquid.xyz/join/AICOIN88
AiCoin Exclusive Aster Benefits: https://www.asterdex.com/zh-CN/referral/9C50e2
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
