Gigawatt AI Plans Turn the Compute Race Into an Execution Test
ByteDance’s prospective Inner Mongolia expansion and Broadcom’s projected customer deployments show how AI infrastructure ambitions are increasingly measured not only in chips, but in power, water, financing and build-out capacity.
By Calder Rowe · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess a real career history, sources, interviews, or firsthand experience.
Key points
- ByteDance is in preliminary discussions to add major computing capacity in Ulanqab, while Broadcom describes prospective large-scale deployments of custom AI processors for leading AI labs.
- The two developments describe different points in the infrastructure chain: ByteDance’s reported plan concerns a physical data-centre cluster, whereas Broadcom’s projections concern processor shipments and customer deployments.
- Ulanqab’s appeal includes renewable energy and cooler conditions, but local officials have also identified scarce water resources and called for tighter controls and water-saving technology in new cloud-computing projects.
Sources: S1
- The most consequential milestone is not a stated capacity target. It is whether sites, grid supply, cooling systems, chips and financing arrive together in a usable operating facility.
Capacity claims now span the whole AI stack
The latest wave of AI infrastructure claims is making the unit of competition more concrete. It is no longer enough to say that a company will buy accelerators or improve a model. The relevant commitments now reach from chip design and manufacturing through financing, power procurement, sites, cooling and operations. ByteDance’s reported effort to enlarge a data-centre cluster in Ulanqab and Broadcom’s outlook for deployments by AI-lab customers illustrate that shift from abstract compute demand to infrastructure that has to be physically delivered.
The reports should not be collapsed into a single project or treated as equivalent evidence. ByteDance’s plan was described by a person familiar with the matter as preliminary discussions with local vendors. Broadcom’s figures came from its chief executive’s account of expected customer deployments and the company’s commercial outlook. One is a reported site-development intention; the other is a supplier’s view of prospective demand for processors. Both point to scale, but each carries different execution dependencies and a different degree of confirmation.
Ulanqab is becoming a concentrated compute location
The reported ByteDance expansion would place additional capacity in Ulanqab’s Jining district, an area already drawing investment or planned facilities from a range of technology companies. Local government and media reports say ByteDance already has operations in the city, and its Volcano Engine unit signed an agreement with the municipal government to establish a computing facility there. Officials have also said local AI data centres support activities including model training and autonomous-driving simulation for companies that include ByteDance, Huawei and Alibaba.
Sources: S1
This concentration has a practical logic. The report attributes Ulanqab’s rise as a northern China data-centre hub to relatively inexpensive renewable energy and cooler temperatures that can lower cooling costs. Those conditions matter because dense AI equipment produces substantial heat and requires continuous electricity. But a favorable starting environment is not the same as unconstrained capacity. A cluster must still absorb rapid construction, connect reliably to power, secure equipment, manage cooling, and operate within local resource limits.
Sources: S1
Sources: S1
Chip road maps meet site realities
Broadcom’s presentation supplies the upstream counterpart to these site ambitions. Its chief executive said Anthropic was expected to deploy TPU 8i chips at a large scale and that Broadcom had visibility toward additional capacity. He also cited a planned deployment of the Jalapeno chip for OpenAI, alongside a broader potential deployment for that program and its next generation. The company described continuing shipments to OpenAI and expected production shipments of Meta’s custom accelerator, while also accelerating shipments of Google processors to Anthropic and Google.
Sources: S2
These statements show why processor supply cannot be assessed in isolation. A shipment forecast becomes economically meaningful only when a customer can install and power the hardware, network it, cool it, and keep it running. Conversely, a data-centre announcement does not create usable AI capacity if accelerators, servers or related infrastructure arrive late. The central risk in an AI build-out is therefore coordination failure across layers that have traditionally been planned separately: silicon supply, capital structure, utility access, facility construction and day-to-day operations.
Financing is part of the infrastructure
The financial burden implied by these plans is substantial even before operations begin. The ByteDance report cites a securities research estimate for upfront investment in a one-gigawatt AI data centre and says a larger proposed facility could entail a much larger outlay. It also reports that ByteDance had secured a large bank loan, according to Bloomberg. ByteDance did not respond to the publication’s request for comment on the reported data-centre discussions.
Sources: S1
Broadcom, for its part, said it could provide residual-value guarantees that would be contingent liabilities to AI-lab customers. Its finance chief framed this as a way to help strategic customers bridge the gap between their existing cash flow and the upfront investments their businesses require. That is a notable signal: the financing challenge may extend beyond customers simply purchasing chips. Suppliers may become more exposed to whether customers can fund and deploy infrastructure at the rate envisioned in commercial forecasts.
Sources: S2
Water turns location advantage into a constraint
The sharpest local constraint identified in the ByteDance report is water. It says data centres can consume significant water for cooling, especially when increasingly powerful AI equipment creates more heat. Ulanqab’s government says water resources are relatively scarce and has called for tighter control of water consumption and the use of water-saving technologies in new cloud-computing projects.
Sources: S1
This does not establish that any individual facility will fail to obtain sufficient water or that every cooling design has the same resource profile. It does establish that announced compute capacity must be judged against more than land and power availability. In a rapidly expanding cluster, water policy, cooling technology and enforcement can determine whether a facility’s theoretical electrical capacity can be sustained in operation. A location’s cooler climate may reduce cooling needs, but it does not remove the need to demonstrate workable resource management.
Sources: S1
Sources: S1
The grid is the shared bottleneck
Neither report provides a completed grid-delivery schedule for the projects and deployments it describes. That absence is consequential. Gigawatt-scale language is a statement about an exceptionally large load, not proof that transmission, interconnection, substations and generation are ready at the required time. Ulanqab’s renewable-energy access is an advantage cited in the report, but a useful AI campus needs dependable electricity where and when its hardware is installed.
For suppliers, the same issue affects the quality of demand. Broadcom can have customer interest and an expanding product pipeline, yet the realized pace of processor deployment will depend on customers’ ability to commission facilities. For operators, it means a chip procurement or financing arrangement cannot substitute for power delivery. Announcements can therefore advance on different clocks: commercial agreements may move first, silicon later, site construction after that, and usable compute only when all of those pieces converge.
Delivered capacity needs a higher evidentiary bar
For ByteDance, delivered capacity would be more than preliminary vendor talks or a government agreement involving its cloud unit. It would mean a built and operating facility with the power, cooling and water arrangements necessary to run AI workloads. The report says the company was aiming for delivery by early 2028, but the expansion details originated with an unnamed person familiar with the matter and ByteDance did not comment. That makes the target a reported intention rather than a confirmed completion timetable.
Sources: S1
For Broadcom, delivery would be more than a forecast of AI revenue or an executive’s stated line of sight to customer deployments. It would be processors produced, shipped, installed and operating in customers’ facilities. Broadcom reported strong semiconductor revenue and laid out expectations for future AI revenue, but those outcomes remain dependent on its customers’ programs and infrastructure readiness. The distinction matters because a supplier can recognize demand before an operator has transformed that demand into durable computing service.
Sources: S2
What to watch next
The next useful evidence from Ulanqab would be more concrete than capacity aspirations: confirmed construction, disclosed power arrangements, cooling design, water-use controls and signs that local policies can accommodate expansion without weakening resource safeguards. It will also matter whether the concentration of projects from ByteDance, DeepSeek, RedNote and other companies creates a durable ecosystem or exposes shared constraints. The existing cluster can support scale, but it can also make competition for infrastructure more acute.
Sources: S1
For Broadcom and the AI labs it serves, investors and policymakers should watch whether projected deployments translate into installation milestones and whether financing tools remain limited or become a broader feature of infrastructure procurement. Broadcom’s outlook demonstrates how strongly chip suppliers are positioning around custom silicon for AI labs. Yet the wider system effect will be determined by construction and utility capacity as much as by processor design. The AI race is increasingly an industrial delivery problem, and its most important proof will be operating infrastructure rather than announced ambition.
Why it matters
Large AI-compute targets can shape chip orders, lending and local development plans long before the underlying facilities are ready. The ByteDance and Broadcom reports show that the limiting factor may be the ability to coordinate finance, hardware, grid access, cooling and water stewardship. Treating capacity claims as delivered capacity too early risks overstating both near-term AI supply and the reliability of the infrastructure supporting it.
Sources
- Exclusive | ByteDance to expand massive AI data centre cluster in Inner Mongolia, source says — South China Morning Post · China Tech ·
- Broadcom delivers strong earnings view as CEO touts growth with AI labs — CNBC Technology ·