Compute & Infrastructure
Top Line
SpaceX is planning a $55 billion 'Terafab' chip fabrication plant in Austin, Texas — if realised, this would represent one of the largest single semiconductor manufacturing investments in US history and a direct attempt to create a vertically integrated AI compute stack outside the TSMC-NVIDIA duopoly.
TSMC reported 17.5% sales growth in April, confirming that hyperscaler AI capex remains robust and that the extended buildout cycle is not yet showing signs of demand saturation.
Nvidia is investing up to $2.1 billion directly into data centre developer IREN, signalling a strategic shift from pure chip sales toward owning or partnering in the physical infrastructure stack that runs its hardware.
An 18-day strike by Samsung chip workers — who rejected a $340,000 one-time bonus in favour of recurring payouts like those at SK Hynix — could cost Samsung up to $11.7 billion and further stress HBM and logic chip supply chains already under pressure.
Michigan towns are rushing to pass blocking legislation after a $16 billion Stargate data centre overrode local opposition, with organised community and bipartisan political resistance now emerging as a structurally significant constraint on US data centre buildout.
Key Developments
SpaceX's Terafab Bet: A $55 Billion Challenge to the Semiconductor Establishment
SpaceX has filed a public hearing notice in Grimes County related to a planned 'Terafab' chip manufacturing facility in Austin, Texas, with a capital commitment of at least $55 billion, according to reporting by The Verge, citing the New York Times and CNBC. This remains at the announced-plan stage — no groundbreaking has been confirmed, and semiconductor fabrication at leading-edge nodes requires years of construction and tool installation before first wafers ship. The investment figure is nonetheless significant: TSMC's Arizona fabs represent a combined commitment of roughly $65 billion across multiple phases, suggesting Terafab is being scoped as a comparable-scale endeavour rather than a niche pilot.
The strategic logic is vertical integration. Musk's xAI already operates a 220,000-GPU supercluster (Colossus) and has separately agreed to provide compute to Anthropic, as confirmed by Tom's Hardware. Controlling fab capacity would close the loop from silicon to inference. The critical unanswered question is process node: no technology partner or EUV licensing agreement has been announced, and building advanced logic fabs without ASML extreme ultraviolet tooling and TSMC process know-how is not a near-term possibility. Whether Terafab targets leading-edge logic, advanced packaging, or a more achievable node remains unconfirmed.
TSMC's 17.5% Revenue Growth Confirms AI Buildout Cycle Has Not Peaked
TSMC reported April sales growth of 17.5% year-on-year, driven by sustained hyperscaler spending on AI training and inference hardware, according to Bloomberg. This follows a pattern of consistent double-digit growth quarters and directly contradicts narratives of imminent capex pullback from cloud providers. South Korea's current account surplus simultaneously hit a record high in March on surging semiconductor exports, per Bloomberg, reflecting strong demand flowing through Samsung and SK Hynix for HBM and NAND alongside TSMC's logic leadership.
The demand signal is reinforced by Nvidia's $2.1 billion equity investment in data centre operator IREN, reported by Bloomberg, framed as accelerating AI infrastructure construction. Nvidia taking equity stakes in compute operators — rather than simply selling GPUs — reflects confidence that demand will remain structurally elevated and that securing physical deployment capacity is now a competitive priority. Combined with Principal Financial Group seeking $3 billion across two data centre funds for US and European deployment, per Bloomberg, capital continues to flow toward physical AI infrastructure with no visible inflection point.
Samsung Strike Risk Adds Labour Disruption to HBM Supply Chain Vulnerabilities
Negotiations between Samsung management and the National Samsung Electronics Union have broken down, with workers rejecting a one-time bonus of approximately $340,000 and demanding recurring annual payouts comparable to the roughly $900,000 distributed to SK Hynix employees following that company's AI windfall, according to Tom's Hardware. An 18-day strike is now threatened, with estimated costs to Samsung of up to $11.7 billion.
The timing is structurally significant. Samsung is already under competitive pressure from SK Hynix in HBM3E supply — SK Hynix dominates Nvidia's HBM allocation — and any production disruption would further delay Samsung's effort to qualify its HBM3E product with Nvidia and recapture share. Samsung's foundry division, which competes with TSMC for advanced logic orders, would also be affected. A prolonged strike would not create immediate system-wide chip shortages given inventory buffers, but would reinforce SK Hynix's HBM advantage at precisely the moment Samsung needs execution discipline most.
Energy and Local Opposition Emerge as Structural Constraints on US Data Centre Buildout
Michigan communities are now mounting organised resistance to data centre construction following the approval of a $16 billion Stargate facility that will consume 1.4 gigawatts of power — approved despite local opposition — with responses including county resolutions, bipartisan state legislation, and a regional water authority refusing utility service to proposed facilities, according to Tom's Hardware. This pattern is spreading beyond Michigan and represents a qualitative shift from isolated NIMBY objections to coordinated legislative and regulatory strategies.
The energy constraint is reinforced by Microsoft's ongoing effort to restart Three Mile Island nuclear plant to power its AI workloads, detailed by Bloomberg. The fact that Microsoft is accepting the reputational and regulatory complexity of nuclear reactivation — one of the most fraught infrastructure projects in modern US history — is an indicator of how constrained grid-connected power procurement has become for hyperscale operators. ARM CEO Rene Haas separately flagged growing data centre demand as the key driver of Arm's business in a Bloomberg interview, underscoring that power availability, not chip supply, is increasingly the binding constraint on deployment velocity.
AMD's MI350P and Baidu's Kunlunxin IPO Signal Broadening Competition at the Hardware Layer
AMD has launched the Instinct MI350P, a PCIe form-factor AI accelerator carrying 144GB of HBM3E and based on the CDNA 4 architecture, claiming roughly 40% faster FP16 and FP8 theoretical throughput than Nvidia's H200 NVL, according to Tom's Hardware and confirmed by ServeTheHome. The PCIe form factor is strategically important: it is a drop-in upgrade for existing air-cooled enterprise servers, removing the infrastructure barrier that has slowed MI300X adoption relative to Nvidia's NVLink-based systems. This addresses a specific enterprise segment where power and cooling constraints preclude liquid-cooled NVL72 rack deployments.
Simultaneously, Baidu's chip unit Kunlunxin is planning dual IPO listings on Shanghai's STAR Market and Hong Kong, per Bloomberg, seeking to capitalise on investor appetite for domestic Chinese semiconductor exposure. Kunlunxin produces AI inference chips and represents China's strategy of developing indigeneous accelerator supply chains insulated from US export controls. A successful dual listing would give Kunlunxin independent capital to scale production, reducing its dependence on Baidu's balance sheet.
Signals & Trends
The Consumer Semiconductor Market Is Being Deliberately Sacrificed to Feed AI Capacity
Motherboard sales across Asus, Gigabyte, MSI, and ASRock are projected to fall by more than 28% in 2026, with the industry selling approximately 11.7 million fewer units, according to Tom's Hardware, explicitly because chipmakers are redirecting wafer and substrate capacity toward AI accelerator production. This is not a demand-side collapse — it reflects a deliberate allocation decision by TSMC, substrate suppliers, and component manufacturers to prioritise higher-margin AI orders. The signal for infrastructure analysts is that the AI buildout is now large enough to create visible supply scarcity and price inflation in adjacent consumer markets, which will generate political and regulatory pressure as PC costs rise. It also means that any AI capex slowdown would release significant manufacturing capacity back into consumer channels, creating a potential inventory correction similar to the 2022 crypto/PC demand crash.
Nvidia Is Transitioning from Chip Vendor to Vertically Integrated Infrastructure Stakeholder
Nvidia's $2.1 billion equity investment in IREN follows a pattern of direct infrastructure involvement — including its own DGX Cloud arrangements and partnerships with neocloud operators — that signals the company is positioning itself to capture economic value across the entire AI compute stack, not just at the silicon layer. This mirrors what hyperscalers did when they moved from server buyers to custom silicon designers. The risk for the ecosystem is that Nvidia as both chip supplier and infrastructure investor creates conflicts of interest in hardware allocation: operators with Nvidia equity relationships may receive preferential GPU supply during constrained periods. Infrastructure buyers without such relationships should treat this as a structural procurement risk and accelerate AMD, Gaudi, and custom silicon qualification programmes accordingly.
Nuclear Power Is Becoming a Credible, Not Merely Aspirational, Component of AI Energy Strategy
Microsoft's Three Mile Island restart is progressing as a confirmed operational plan rather than a speculative announcement, representing the most tangible evidence yet that hyperscalers are willing to accept extraordinary complexity and regulatory exposure to secure large blocks of carbon-free baseload power. The precedent matters beyond Microsoft: Google has contracted with Kairos Power for small modular reactors, and Amazon has made similar commitments. As grid interconnection queues in primary data centre markets stretch to 5-7 years, nuclear — both large plant restarts and SMRs — is transitioning from an experimental hedge to a mainstream procurement option for operators who need more than 100 megawatts of firm power. Infrastructure investors and site selectors should weight proximity to operational or restartable nuclear capacity as a first-order site selection variable, not a secondary consideration.
Explore Other Categories
Read detailed analysis in other strategic domains