Third-Party Chip Manufacturers Enter the 'Countdown'

To keep profits in-house, ByteDance chooses to develop its own chips.

Recently, according to BlueWhale Tech, ByteDance's chip R&D team will begin large-scale recruitment, with hiring taking place in cities such as Beijing, Shanghai, and Shenzhen. Positions include chip architecture, SoC design, and more.

Insiders reveal that ByteDance's chip R&D team is currently focused on chip design, developing custom hardware and optimizations tailored to the company's own business needs. They are researching and developing multiple complex chips using advanced semiconductor processes for cloud-based scenarios, aiming to improve performance and reduce computing power costs.

'Currently, ByteDance's chip team has achieved multiple successful first-pass silicon runs, with several early projects entering mass production deployment stages. These cover multiple mainstream advanced process nodes, and the overall R&D and implementation progress is steady.'

In terms of team size, according to Radar Finance, ByteDance's AI chip team and CPU team are relatively large, with the AI chip direction having over 500 members and the CPU team around 200.

However, the assembly of a thousand-person team is just the beginning—the real competition is only now unfolding.

Faced with growing AI investments, can self-developed chips help ByteDance completely escape the curse of 'working for NVIDIA' and transform expensive computing power costs into exclusive profits for Volcano Engine? Can it replicate Google's TPU success story, using extreme software-hardware synergy to redefine pricing power in the MaaS market?

For domestically produced chip manufacturers like Cambrian, which have only just become profitable, does each successful tape-out by major players signal market prosperity or the cruel 'countdown to disintermediation'?

01

The Expensive 'Toll'

The Hundred-Billion-Yuan Cost ByteDance Must Consider

Developing its own chips is an economic imperative for ByteDance.

According to the Financial Times, ByteDance's planned capital expenditures for 2026 are set to reach approximately RMB 160 billion, up from around RMB 150 billion in 2025. Roughly half of this funding will be used to purchase advanced semiconductor chips for developing AI models and applications. If approved to purchase more advanced chips, capital expenditures will increase significantly further.

If we rewind five years, ByteDance could afford and was willing to spend on AI chips.

At that time, AI was merely an 'optimization item' in ByteDance's portfolio, primarily used to improve recommendation algorithms or enhance the precision of advertising systems. However, with the explosive growth of large models like 'Doubao' and AI applications such as Doubao and Jimeng, the situation has fundamentally changed. According to QuestMobile data, the Doubao app had the highest monthly active users among domestic AI applications by the end of 2025, with user engagement growing exponentially.

Meanwhile, Volcano Engine, ByteDance's new growth engine for B-end business, is experiencing unprecedented demand for computing power.

After initiating an aggressive price war that brought large model pricing into the 'fen era,' Volcano Engine is now processing hundreds of trillions of tokens daily through its extreme cost-performance strategy. While this price-for-volume approach has captured market share in MaaS, it has also pushed the underlying computing power to its breaking point. Volcano must not only support real-time interactions for hundreds of millions of internal users but also respond to massive API calls from external enterprise clients.

Thus, as AI business has shifted from a 'value-add' to a 'core growth engine,' the nature of computing power consumption has changed: it is no longer an R&D investment but an operational cost that bleeds money every minute.

As tens of thousands of GPUs roar in data centers, every second that passes is not just an electricity cost but also a massive profit margin being sliced away by upstream hardware vendors like NVIDIA. For 'Doubao' and video generation models with hundreds of millions of daily active users, the computing power consumption during inference far exceeds that of training—precisely the 'disaster zone' for cost control.

Continuing to rely on expensive general-purpose GPUs to process vast user requests means not only continuously 'working for suppliers' but also struggling to gain pricing initiative when providing MaaS services externally.

This is no exaggeration—just look at the suppliers' financial reports to see how brutal this 'employment' is.

In the seller's market for AI chips, pricing power rests entirely with upstream players. According to The Information, NVIDIA's gross margins on high-end GPUs have long remained above 70%. For a super-large client like ByteDance, which makes hundred-billion-yuan purchases, this means that for every RMB 100 spent on chips, RMB 70 goes toward paying Jensen Huang's 'tolls.'

Even in the domestic chip sector, seen as an 'affordable alternative,' the situation hasn't improved much. Take Cambrian, one of ByteDance's core suppliers, as an example. Its 2025 financial report shows that benefiting from frantic purchases by internet giants, the company not only achieved historic profitability but also saw margins for its ThinkForce series inference cards climb due to economies of scale, with overall gross margins exceeding 50%.

Thus, a harsh but clear logic emerges: the excess profits of upstream chip manufacturers are essentially cost sinkholes for ByteDance and Volcano Engine.

In an era of stock competition (stock competition), every yuan of supplier profit represents cost-saving potential for ByteDance. Since inference chips don't require the extreme generality and double-precision computing power of training chips, why surrender hundreds of billions in profits to others?

Moreover, in the global tech giant arena, 'self-developed chips' are no longer a cost-related multiple-choice question but a proof-of-capability exam—proving whether you qualify to be a true cloud giant.

Looking overseas, Google began developing TPUs thirteen years ago. While the world scrambled for NVIDIA GPUs, Google quietly rolled out TPUs. Through the software-hardware integration of TPUs and the TensorFlow framework, Google met the massive computing power demands of its search, YouTube recommendations, and other services while slashing unit computing costs to a fraction of competitors'—even beginning to sell TPUs externally, acting as a 'second NVIDIA.'

Back home, Alibaba Cloud has also taken the self-development route.

In late January, T-Head Semiconductor launched its latest-generation AI chip, 'Zhenwu 810E,' directly competing with NVIDIA's H20 in performance. With 100,000 units shipped, it surpasses Cambrian and ranks among the top tier of domestic GPU manufacturers. Deployed in multiple 10,000-card clusters within Alibaba Cloud, it serves over 400 clients, including State Grid and XPENG Motors.

More importantly, Alibaba Cloud has achieved a closed loop with the 'T-Head + Alibaba Cloud + Tongyi Qianwen Model' golden triangle.

According to Zhang Tao, an inference framework optimization expert interviewed by Lei Feng Network, 'Take DeepSeek's launch of a large MoE model as an example—its essence is achieving extreme model-hardware co-design in cloud-based cluster scenarios to maximize computing performance. Alibaba's full-stack self-development represents an even more 'aggressive' native adaptation approach.'

For ByteDance, with Google's TPU as a 'shining example' and Alibaba's PPU as a 'close demonstration,' developing its own chips now is not just defensive cost control but an offensive move to replicate Google's success in the MaaS war. When computing costs determine the floor for model pricing, cloud vendors without self-developed chips are doomed to be NVIDIA's 'ticket sellers,' forever unable to touch the core profits of the business closed loop (closed loop).

To avoid this fate, ByteDance must master 'tailored' technology.

As Zhang Tao puts it, extreme co-design is the key to breaking through for internet companies. Compared to purchasing general-purpose GPUs, self-developed ASIC chips are like precision scalpels, deeply customized for ByteDance's unique recommendation algorithms and Transformer architectures. They strip away all redundant general-purpose functions, ensuring every transistor is used effectively.

Industry estimates suggest this extreme specialization can reduce unit computing costs by over 50%. This is not just a technical victory but a qualitative shift in business logic: once procurement scales surpass the 'self-development break-even point,' developing chips transitions from a nice-to-have technical reserve into a survival threshold that giants must cross in competition.

Only by fully controlling the 'computing power faucet' can ByteDance face future brutal model price wars with the confidence of not needing to please others.

02

Chip 'Disintermediation'

Who Is Shivering in the Cold?

However, what is sweet for one party is often poison for another in the business world.

When ByteDance decides to keep those hundreds of billions in 'excess profits' in-house, for domestic chip manufacturers that only just tasted success in 2025, this means not just losing a major client but the start of a life-and-death countdown to 'disintermediation.'

Recalling the origin of this relationship, ByteDance's partnership with Cambrian was always a 'spare tire' from the start.

Two years ago, this was a 'huddling for warmth' under extreme external pressure. Restricted by export bans, A100/H100 supplies were cut off. Although 'neutered' versions like H800 and H20 were later introduced, the Sword of Damocles hung over every internet giant's head. During the darkest days of the supply chain, ByteDance desperately needed a 'Plan B' to hedge against NVIDIA's risk of sudden supply cuts.

As the 'first domestic AI chip stock,' Cambrian, with its ThinkForce series, became one of the few suppliers with Mass production spot goods (mass-produced inventory) and a relatively mature software stack.

This was a classic 'fear-based purchase,' not a 'performance-based preference.'

According to industry chain research, between 2023 and 2024, while ByteDance was frantically purchasing NVIDIA-compliant chips, it also placed sizable Indent Order (letters of intent) with Cambrian. These chips weren't immediately deployed in core recommendation algorithm clusters but were used for testing inference scenarios in non-core businesses.

This cooperation peaked in 2025: as ByteDance's AI products exploded, inference computing demand surged, and NVIDIA's capacity tightened, Cambrian's ThinkForce 590 successfully entered ByteDance's supply chain, filling part of the computing gap.

Thanks to this 'risk-averse purchasing' by ByteDance, Baidu, Alibaba, and other internet giants, Cambrian reached its brightest moment since founding.

In 2025, Cambrian's revenue doubled to over RMB 6 billion, a year-on-year growth rate exceeding 200%. More importantly, it crossed the financial chasm that had plagued it for years, achieving its first-ever annual profit.

In capital markets, Cambrian's stock price soared on this 'better-than-expected' performance. From early 2025, its stock price rocketed, breaking through major resistance levels, with its market cap once nearing RMB 500 billion. During that period, brokerage reports shouted in the most excited tones that 'China's NVIDIA has been born,' tech stock forums discussed how the ThinkForce series would become ByteDance's 'crown prince,' and even rumors emerged about ByteDance acquiring a stake in Cambrian.

Investors firmly believed that the endgame of domestic substitution had arrived—Cambrian would lie in the computing power arms race of internet giants like NVIDIA did, enjoying high-margin dividends.

However, beneath this prosperous facade lay an extremely fragile foundation.

Cambrian's explosive growth was essentially built on two temporal gaps: 'NVIDIA supply shortages' and 'delayed giant self-development.' This was a classic 'window-period dividend.' For ByteDance, large-scale purchases from 2024 to 2025 were more a 'fear-based hedge' and 'mandatory task' than a technology- and trust-based cooperation.

Industry chain insiders revealed to Super Focus that ByteDance's core demand for purchasing Cambrian's ThinkForce 590 chips was simple: fill gaps. During the window when NVIDIA's high-end computing power was restricted and domestic self-developed chips weren't yet mass-produced, ByteDance needed massive inventory to sustain the exponential growth in inference demand for the 'Doubao' model. As the largest internet company, it also had a responsibility to support domestic computing power.

In other words, giants bought from you not because you were irreplaceable but because you were the only one 'in stock' at the time.

By 2026, when ByteDance's 1,000-person chip team completes the full puzzle from architecture design to tape-out verification, this expensive 'spare tire' will gradually lose its purpose.

As an independent chip design company, Cambrian's product iteration cycles typically follow traditional hardware patterns—a chip takes 18 to 24 months from definition to mass production. This means when ByteDance's algorithm team proposes a new sparse attention mechanism in 2026, using Cambrian's chips might delay hardware acceleration until late 2027.

But internally at ByteDance, this logic is completely rewritten.

With the 'SeedChip' project land (implemented), ByteDance's algorithm team can sit in the same office as the chip team. This zero physical distance creates a chemical reaction: an operator optimization demand proposed by algorithm engineers in the morning can be written into the instruction set by chip architects in the afternoon, or even directly adjust on-chip memory ratios. Through this extreme co-design, ByteDance can achieve hardware-level iteration support within months.

When ByteDance can meet its needs at lower costs and faster iteration speeds, continuing high-margin purchases from Cambrian violates basic business logic.

This crisis belongs not just to Cambrian.

Broadening the view to the entire domestic chip sector, unicorn companies like Biren Technology and Moore Threads face the same risk of being 'dimensionally crushed' by giants. Their business models largely aim to become 'China's NVIDIA' by designing general-purpose high-performance GPUs for cloud vendors and smart computing centers.

But the reality is that their largest potential clients—Alibaba Cloud, ByteDance, Tencent Cloud, and Baidu Intelligent Cloud—are all aggressively pursuing self-developed chips.

Alibaba's PPU has already proven this path's feasibility. Its self-developed AI chips are widely deployed in Alibaba's internal recommendation and inference businesses and are exploring external sales. If Alibaba Cloud was the first to eat this crab, ByteDance is the predator following closely behind with an even bigger appetite.

This is the fate of 'middlemen' in the era of giant competition.

In the cloud and AI industrial chain, midstream third-party independent chip designers face unprecedented 'sandwich extrude (squeeze).'

Upward, they lack NVIDIA's impenetrable CUDA ecosystem moat, which makes clients 'have to buy.' Downward, they face extreme cost-performance competition from giant-developed chips. Thus, when the largest client becomes the strongest competitor, third-party vendors' survival space shrinks to the limit.

In the future, internet giants like ByteDance and Alibaba will never (definitely not) confine self-developed chips to their own data centers as 'internal exclusives.' Business logic always pursues maximum economies of scale: once 'SeedChip' or 'Zhenwu' has been polished and validated across millions of cards internally, with stable yields and extremely low costs, they will inevitably move toward external sales.

By then, these Internet giants will have accomplished the most formidable crossover: they will be both the largest cloud service providers and the largest chip suppliers. Under such a dimensionality reduction assault, only two paths remain for 'specialized survivors' like Cambrian.

The first path is to become the next NVIDIA.

To tread this path, you must do more than just design the best circuit diagrams; you must build an impregnable software ecosystem moat like CUDA. Globally, NVIDIA stands alone as the independent vendor that has navigated this path successfully, requiring an Ultimate cooperation [Note: This Chinese phrase is translated literally as 'extreme alignment of heaven, earth, and humanity' but could be contextually rendered as 'perfect alignment of timing, resources, and team' if more background is provided] of 'timing, location, and teamwork.'

Looking back at NVIDIA's rise, it was built on a 15-year 'lone marathon.' Jensen Huang burned hundreds of millions of dollars annually to promote CUDA in the pre-AI wilderness. Back then, no giants watched him, no cloud vendors poached his business, and he had ample time to 'grow stealthily,' transforming graphics cards from gaming tools into AI engines bit by bit.

But now, that 'timing' has vanished completely.

Today, with AI at the core of the global tech arms race, any startup attempting to replicate CUDA's path will find giants already lurking everywhere. You won’t have a decade to nurture a developer ecosystem because ByteDance and Alibaba’s self-developed chips will launch tomorrow, bypassing your ecological barriers with direct PyTorch/TensorFlow Bottom layer adaptation [Note: 'bottom-layer adaptation']. You won’t have room for trial and error because capital markets and clients no longer believe in long-cycle myths.

This path means rebuilding NVIDIA under the giants’ noses—a brutal breakthrough [Note: 'breakthrough' or 'do-or-die effort'] with slim survival odds.

The second path, however, is a more pragmatic 'sell-out' to the Internet giants.

For most domestic chip vendors unable to become NVIDIA, this might be the best outcome. Rather than being slowly strangled by 'disintermediation' in the giants’ cracks, they could choose to merge into the giants’ empires while their technology still holds value and their teams remain combat-ready, integrating as a hardware division within ByteDance or Alibaba. This isn’t failure but another form of 'landing safely.'

For Cambrian, profitability in 2025 might mark a brief glorious dawn—or the last hurrah before the giants' 'disintermediation.'

In this winner-takes-all era of computing power, pure 'hardware middlemen' are losing their foothold. On the future poker table, only two types of players remain: those strong enough to set rules like NVIDIA, making giants work for you, or those smart enough to integrate into giants like Annapurna, becoming part of the rules themselves.

No other easy paths remain.

- END -