ByteDance Bets Big on Northern China for Next-Wave AI Compute

Racing to Scale Compute in the North

ByteDance is negotiating to add between five and six gigawatts of computing capacity in Ulanqab, a city in Inner Mongolia that has quietly become one of China's most important AI infrastructure hubs. The timeline spans the next two years, and the expansion would represent one of the largest single commitments to data centre capacity by a Chinese internet company in recent memory. At DailyTechWire, we have tracked the steady migration of hyperscale workloads from coastal provinces to the interior, driven by land costs, power availability, and provincial incentives. This move by the TikTok parent fits squarely into that pattern, but the scale is striking even by the standards of China's infrastructure boom.

Inner Mongolia has emerged as a favored destination for compute-hungry companies because of its proximity to coal-fired and renewable power sources, relatively cool climate that lowers cooling costs, and land availability that coastal tier-one cities can no longer offer. ByteDance already operates data centres in the region, and the new capacity would consolidate Ulanqab as a primary node in the company's AI training and inference network. The negotiations involve multiple local data centre vendors, reflecting the complexity of assembling gigawatt-scale power and cooling infrastructure in a landlocked region where grid reliability and water access remain variables.

Why Inner Mongolia, Why Now

The choice of location and the timing both reflect strategic pressures. ByteDance has been racing to build out its large language model capabilities, particularly for its Doubao family of models, which power conversational features across its apps and serve enterprise customers through its Volcano Engine cloud platform. Training and fine-tuning these models at scale demands thousands of GPUs running in parallel, which in turn requires stable, low-latency interconnects and enormous amounts of electricity. Ulanqab offers all three at a price point that Beijing or Shanghai cannot match.

At the same time, U.S. export controls on advanced AI chips have forced Chinese companies to rethink their hardware strategies. Nvidia's H100 and successor architectures are effectively off-limits, which means ByteDance and its peers are relying on a mix of older Nvidia silicon, domestically produced alternatives from vendors such as Huawei and Moore Threads, and custom ASICs where workloads allow. None of these options can match the performance density of the latest U.S. chips, so the compensating strategy is to deploy more of them. That arithmetic explains why Chinese tech giants are building out capacity at a pace that might otherwise seem extravagant.

The funding rounds we have followed across the region suggest that ByteDance is far from alone. Alibaba, Tencent, and Baidu have all announced multi-billion-dollar data centre expansions over the past eighteen months, many of them concentrated in Inner Mongolia, Ningxia, and Gansu. Provincial governments in these regions have offered tax breaks, subsidized land, and priority access to grid capacity in exchange for the jobs and prestige that come with hosting a tier-one tech tenant. The result is a kind of infrastructure gold rush, with local officials competing to attract hyperscalers and hyperscalers competing to lock in power and rack space before the next wave of model releases.

The Economics of Gigawatt-Scale AI

Five to six gigawatts is an enormous figure by any standard. For context, a single gigawatt can support roughly one million high-performance servers, depending on power density and cooling architecture. ByteDance's planned expansion would therefore add capacity equivalent to several million CPU cores or hundreds of thousands of GPUs, assuming a mix of training and inference workloads. The capital expenditure required to build, equip, and operate facilities at this scale runs into the billions of dollars, even in a lower-cost region like Ulanqab.

The economics hinge on two variables: utilization and amortization. If ByteDance can keep the new capacity running at high utilization, training models around the clock and serving inference requests during peak hours, the per-query cost of compute drops rapidly. But if demand falls short or if the models being trained prove less commercially successful than anticipated, the company is left with stranded assets and a heavy depreciation burden. This is the bet every hyperscaler makes when it commits to gigawatt-scale infrastructure: that the demand for AI services will grow fast enough to justify the fixed costs.

In ByteDance's case, the bet appears grounded in both internal demand and external revenue. The company's consumer apps, led by Douyin and TikTok, generate recommendation and moderation workloads that already consume vast amounts of compute. Doubao and other generative AI features are layered on top of that base load, and Volcano Engine is positioned to capture enterprise spend from companies that want to build AI applications without managing their own infrastructure. If any Chinese internet company has the revenue base to support a multi-gigawatt expansion, ByteDance is near the top of the list.

Risks and Constraints

Still, the expansion carries risk. Power availability in Inner Mongolia is not unlimited, and the region's grid is heavily reliant on coal. As Beijing pushes for carbon neutrality by 2060, provincial authorities face mounting pressure to curtail emissions, which could translate into higher electricity prices or caps on new data centre approvals. ByteDance and its peers are investing in on-site renewable capacity, including solar and wind, but the intermittency of these sources means they remain supplements rather than replacements for baseload coal power.

Water is another constraint. Data centres at gigawatt scale require millions of liters of water per day for cooling, and Ulanqab sits in a semi-arid region where water rights are contested between industrial users, agriculture, and municipal supply. Local governments have so far prioritized data centre tenants, but that calculus could shift if drought conditions worsen or if public opposition to industrial water use intensifies.

There is also the question of latency. Ulanqab is roughly 400 kilometers from Beijing, which is manageable for batch training workloads but less ideal for real-time inference serving users on the coast. ByteDance will need to architect its network carefully, likely pairing the Inner Mongolia cluster with edge nodes in tier-one cities to keep latency below the thresholds that users notice. This kind of hybrid topology is standard practice for hyperscalers, but it adds complexity and cost.

What It Signals for the Sector

ByteDance's move is a data point in a broader trend: Chinese AI companies are building out capacity at a pace that rivals, and in some cases exceeds, their U.S. counterparts. The constraints imposed by export controls have not slowed the buildout; if anything, they have accelerated it, as companies race to lock in hardware and power before geopolitical conditions shift again. The result is a landscape in which compute, rather than algorithms or data, is becoming the primary competitive moat.

For the rest of the region, the implications are mixed. On one hand, the infrastructure investments create jobs and tax revenue in provinces that have historically struggled to attract high-tech industry. On the other, they lock those provinces into an energy-intensive development path that may prove difficult to reconcile with climate goals. And for the global AI race, the message is clear: China's tech giants are not waiting for permission to build the next generation of models. They are building the infrastructure to do it themselves, one gigawatt at a time.