The material establishes a single central tension for Broadcom Inc.: its custom-silicon franchise, centered on Google TPUs, remains strategically central to AI infrastructure yet faces measurable share dilution as hyperscalers diversify suppliers and foundry capacity stays tight. That distinction matters because the near-term investment case rests less on whether custom silicon grows — demand is treated as a growth input 11 alongside hyperscaler partnership capability 11 and an expanding data center total addressable market cited as an industry tailwind 18 — than on how much of that growth Broadcom retains and can actually ship through a concentrated supply chain.
In Marshallian terms, we must distinguish between the short run, in which design incumbency and allocated foundry capacity determine shipped volume, and the long run, in which customers gradually adjust sourcing, second-source designs, and build new capacity. The interesting question is not whether Broadcom is large in custom silicon, but why that position persists and at what margin it erodes.
The marginal erosion of the TPU franchise
The most direct pressure is the forecast that Broadcom's share of Google TPU revenue declines from 95% in 2026 to 80% in 2027 and 65% by 2028 15. The same 95%-to-80%-to-65% path between 2026 and 2028 is described as MediaTek taking TPU share 15, identifying the challenger at the margin.
The scale at stake is large: a Google Cloud TPU 8t superpod contains 9,600 chips 1,17. Broadcom is said to retain the TPU 8t Superpod business because peak FLOPS still matter for training 17, which preserves relevance in training even as inference diversifies. This is a particularly revealing case: training and inference do not exhibit the same elasticity of substitution across suppliers.
Broader custom-CPU efforts reinforce that diversification risk, with ARM and Meta pursuing parallel custom-CPU efforts 4 and Meta's interest in custom server-class CPUs cited for both Qualcomm and ARM 4. The adjustment is therefore gradual rather than abrupt — nature does not leap — but it extends beyond a single socket or a single customer.
Why hyperscalers second-source: control without replacement
The economics explain why hyperscalers pursue this path despite incumbency. Custom chips require considerable engineering investment 12, but the economic case becomes stronger when they are manufactured and deployed at significant volume 12. At volume, fixed design costs are spread and quasi-rents become attainable.
MediaTek's design capabilities and manufacturing relationships can lower the barrier to developing custom chips 12, while hyperscalers and large AI companies could use custom processors to gain control over roadmaps without replacing every element of their existing Nvidia investment 12. Customers can develop a custom processor for a high-volume workload while preserving compatibility with Nvidia systems 12, and Nvidia says NVLink Fusion can allow custom chips from partners such as MediaTek to communicate with Nvidia processors at high speeds 12.
The purpose of Nvidia's investment in MediaTek is stated as countering Big Tech chip development 10, and Nvidia's rack-level strategy is framed as mitigation against the threat of customer custom-chip development 6. As the processor market becomes more varied, broadly adopted interconnect and systems architecture increases in value 12, which favors vendors that own fabric and system integration, not just silicon.
Foundry concentration and the short-run constraint
That competitive shift collides with persistent foundry concentration. TSMC is described as dominant in advanced logic nodes 3, and TSMC capacity is described as sold out through 2027 or 2028 14, with fabs separately described as sold out through 2027 2. Against that backdrop, TSMC is said to prioritize Nvidia, Apple, AMD, Broadcom, and hyperscalers ahead of Qualcomm 4, a relative advantage for Broadcom in allocation even as absolute capacity binds.
The supply response is large but slow and internally inconsistent in the material: TSMC's expansion is described as including 25 new fabs and packaging facilities worldwide 8, while a separate claim has TSMC building nearly 20 fabrication plants worldwide 9. We must be careful to distinguish here — the disagreement is about the count and scope of the build, not the direction — and it underscores uncertainty about the long-run adjustment path.
Capacity and equipment availability represent the primary strategic constraints in semiconductor competition 9, and constraints are said to have moved upstream from advanced packaging to more basic input materials 7. Execution risk is underscored by Hock Tan's statement that no chip manufacturer can guarantee the pace at which data centers and cloud providers will deploy available capacity 13, even as the growth thesis includes two consecutive doublings supported by the supply chain 13 and the supply chain is expected to support two production doublings if production stays on schedule 13.
Complementary scale signals include AWS's custom-silicon business, including Trainium, Graviton, and Nitro, having separately surpassed a $25 billion annualized run rate disclosed in a July earnings report 16, and MediaTek's $2 billion 2026 ASIC signal and the deployment of 2 million GPUs by AWS identified as scaling indicators 12. The post identifying Hock Tan as the CEO doubling down on supply 5 ties that supply push directly to Broadcom's leadership stance.
What it implies: allocation, partnership, and system-level value
Collectively, the material points to a Broadcom position that is advantaged but less exclusive: incumbency plus prioritized foundry access supports near-term volume, while second-sourcing by the largest customer and Nvidia's compatibility-plus-systems response compress the moat around any single custom chip. Financial leverage therefore hinges on converting data-center TAM growth and custom-silicon ramps into shipped volume through TSMC allocation and upstream materials, rather than on design wins alone. Strategic value migrates toward owning partnership depth with hyperscalers and toward interconnect and system-level integration that remains useful even as training and inference silicon diversifies.
Under current conditions, the evidence suggests three conditional conclusions. First, retaining share in a growing but fragmenting TPU pool is decisive: the 95% to 80% to 65% revenue-share path through 2028 15 makes defending Google volume and extending to additional hyperscaler custom programs the variable at the margin, even as peak-FLOPS training relevance 17 provides a near-term anchor. Second, allocation must be used as a competitive edge: prioritized access alongside Nvidia, Apple, AMD and hyperscalers 4 matters more when leading capacity is sold out into 2027-2028 14 and deployment pace cannot be guaranteed by any chipmaker 13. Third, competition must be met at system level, not chip level: with varied processors raising the value of interconnect and systems architecture 12 and Nvidia countering custom silicon via compatibility and rack-scale integration 6,12, Broadcom's upside depends on pairing custom silicon with fabric, partnership capability 11 and execution of the planned production doublings 13.