The AI ASIC Market, Part 5: The Inference Economy Has Many Winners

TLDR

• NVIDIA isn't losing the AI silicon market. It's repositioning as a full-stack platform via NVLink Fusion, the Groq acquisition, and the Intel partnership.
• Independent AI chip companies collectively represent the heterogeneous inference economy, a structurally different business than fighting NVIDIA on training.
• For infrastructure investors, custom silicon at industrial scale is being priced at generational multiples that dwarf comparable bets in Bitcoin mining infrastructure.
• For Bitcoin miners, the operational playbook already in place — power, cooling, and site economics — translates directly to where AI infrastructure capital is flowing next.

A Quick Recap: Parts 1–4

Parts 1 through 4 mapped the alternatives to NVIDIA. Part 1 covered the hyperscalers building their own silicon to reduce dependency. Part 2 covered the design enablers that capture most of the ASIC services revenue. Parts 3 and 4 covered the independent AI chip companies challenging NVIDIA directly, plus the architectural outliers operating on different strategic logic.

Now let's look at how NVIDIA is responding, because the company at the center of the AI chip market doesn't disappear when alternatives emerge. It adapts.

NVIDIA's Strategic Response

NVIDIA's response to the independent AI chip threat is the most strategically sophisticated in the market. The company isn't competing purely on chip performance. It's repositioning as a full-stack AI platform company, and the moves of the last eighteen months make this clearer than ever.

NVLink Fusion opens NVIDIA's NVLink interconnect to third-party ASICs. Hyperscalers can now embed NVLink in their own chips, getting high-speed connectivity without building their own interconnect. NVIDIA cedes some hardware revenue in exchange for deeper ecosystem lock-in. Fujitsu and Qualcomm have already adopted NVLink Fusion in CPU designs. It's the "give a little to gain more" move. NVIDIA sacrifices some hardware sales to keep customers in the broader NVIDIA software and interconnect ecosystem. When the hyperscalers' custom ASICs need to talk to other chips, they talk through NVIDIA's fabric. That's a durable dependency that outlasts any individual chip cycle.

The Groq acquisition (covered in full above) fills NVIDIA's one structural gap: ultra-low-latency inference on the decode phase of token generation, by acquiring the best independent solution rather than building from scratch. The Groq 3 LPU in Vera Rubin systems targets 35x better tokens per watt for decode-heavy workloads, and NVIDIA's guidance to build data centers with approximately 25% LPU capacity is a careful engineering of how much inference revenue to shift to the new architecture. Not a full replacement. Not a token gesture. A measured integration that captures the architectural advantage Groq's LPU offers while protecting the GPU revenue base.

The NVIDIA-Intel partnership announced in September 2025 for $5 billion extends NVIDIA's ecosystem into x86 CPU territory. Co-development of AI infrastructure and x86 CPUs integrated via NVLink. This is NVIDIA acknowledging that the future data center AI stack isn't just GPUs. It's GPUs plus CPUs plus specialized accelerators plus high-speed interconnect, all tightly integrated.

The broader M&A wave is reshaping the space. AMD acquired Untether AI (RISC-V inference chip). Meta acquired Rivos. Intel invested $350 million in SambaNova. The inference ASIC M&A cycle is in full swing, with NVIDIA's Groq deal setting the price benchmark that every other transaction is now measured against. Cerebras' $35 billion IPO target and Etched's $5 billion private valuation are both, consciously or not, priced against the Groq $20 billion comparable.

The market share math is worth stating clearly. NVIDIA's data center revenue in Q4 FY2026 hit $62.31 billion, up 75% year-over-year. Data center revenue accounts for nearly 88% of total NVIDIA revenue. Gross margins of approximately 73.5% reflect what might fairly be called a monopoly premium. As ASIC alternatives proliferate, analysts expect NVIDIA's AI silicon market share to normalize from the current ~85-86% toward something closer to ~75%. This is gradual, not rapid, given the entrenched CUDA ecosystem advantage. NVIDIA isn't going to lose this market. But it will stop winning all of it.

The Vera Rubin platform is where the full-stack strategy culminates. Seven-chip Rubin SuperPOD strategy detailed at GTC 2026 combines Rubin GPUs, Groq 3 LPUs (from the acquisition), Intel CPUs (from the partnership), and NVLink Fusion interconnect. This is NVIDIA positioning itself not as a GPU vendor but as the orchestrator of the full AI infrastructure stack. If the strategy works, NVIDIA captures the economics of the entire accelerator market even as it cedes specific workloads to specialized silicon. If it doesn't work, NVIDIA's long-term share could erode below the ~75% normalization analysts currently expect.

The Independent AI Chip Landscape: Scorecard

Seven companies, two fundamentally different business models, and valuations ranging from early-stage to $500 billion-plus market caps. The table below summarizes where each stands as of mid-2026.

The read across the table: Broadcom and Marvell are the only public-market pure plays in the independent camp. Groq proved the $20 billion exit thesis. Cerebras is about to prove it at the IPO level. Etched, Tenstorrent, and Tensordyne represent the remaining independent bets at different stages of maturation. The Groq comparable has reshaped every valuation in the sector. Every transaction that happens from here forward will be priced in part relative to that $20 billion benchmark.

Investment Landscape

For public-market investors, Broadcom ($AVGO) and Marvell ($MRVL) are the direct pure plays on AI ASIC design services. Broadcom's scale ($100 billion FY2027 AI revenue target, 78.6% gross margins, $73 billion committed backlog) makes it the more concentrated bet on the hyperscaler ASIC build-out. Marvell offers smaller, lower-valuation exposure with higher customer concentration risk in a secondary design services position. Both benefit structurally from hyperscaler ASIC spend, but Broadcom captures a disproportionate share of the upside.

Cerebras is the only pre-IPO opportunity in this report, targeting a May 2026 Nasdaq listing under the ticker CBRS at a $22-35 billion valuation. Pre-IPO shares have traded at implied $26-28 billion valuations on secondary platforms like Forge and EquityZen, which suggests the IPO may be priced conservatively. Retail investors will have access once CBRS begins trading; accredited investors can access pre-IPO through secondary marketplaces.

Indirect public exposure to hyperscaler AI ASIC strategies runs through the parent companies: Alphabet (Google TPU), Amazon (AWS Trainium), Microsoft (Maia), Meta (MTIA). These aren't pure plays on AI ASICs; they're diversified technology companies where the ASIC program is one component of many.

TSMC ($TSM) is the foundry play. Every AI ASIC (hyperscaler or independent) manufactures at TSMC, or, increasingly, at Samsung for specific products like the Groq 3 LPU. TSMC is the most concentrated supply chain exposure in the entire AI compute stack.

The material risk factors for any AI ASIC investment thesis include:

1. The Warren-Blumenthal antitrust investigation into the NVIDIA-Groq deal, which could create uncertainty across the sector.
2. Customer concentration risk at Broadcom (Alphabet exposure) and Cerebras (historical UAE concentration).
3. Execution risk at pre-silicon companies like Tensordyne and pre-production Etched.
4. Component supply constraints (HBM, advanced packaging, advanced-node foundry capacity) affecting the entire market.
5. Regulatory action on AI model export controls, semiconductor tariffs, or antitrust enforcement that could shift the competitive landscape.

None of this is financial advice. It's a description of the risk factors the market is actively pricing.

What This Means for Bitcoin Miners

The independent AI chip trade has four direct implications for Bitcoin mining operators and the investors who allocate capital to them.

The valuation signal is the first. The entire independent AI chip sector is being validated at numbers that dwarf Bitcoin mining infrastructure companies. Cerebras at $22-35 billion for a business doing $510 million in revenue. Broadcom's AI revenue alone approaching a $50 billion annual run rate. Groq's $20 billion exit on pre-acquisition revenue of roughly $500 million. Every one of these valuations rests on the same industrial thesis that drives Bitmain, MicroBT, and Canaan: custom silicon at industrial scale for well-defined workloads. The market is pricing that thesis at generational multiples in AI; it's pricing the equivalent thesis at mid-cycle multiples in Bitcoin mining. That disparity is either durable (because AI compute genuinely is a larger market) or a signal of misallocation that could correct either direction over time.

The capital availability signal is the second. The AI chip sector is absorbing tens of billions in private and public capital annually. Series H rounds priced at $23 billion. IPOs targeting $35 billion. $20 billion acquisitions. If Bitcoin mining infrastructure companies can credibly articulate their position in the broader custom silicon and compute infrastructure thesis (and miners increasingly can, given overlapping power, cooling, and site economics), there is significantly more capital available than there was in 2022-2023. The AI infrastructure capital flywheel will not leave adjacent industrial compute opportunities untouched indefinitely.

The competitive signal is the third. The independent AI chip companies are eating into NVIDIA's margins in exactly the way Bitcoin mining operators once ate into GPU mining. Custom silicon for specific workloads always wins at scale. This is the same lesson, taught in two different markets, arriving at the same conclusion. For anyone building or funding compute infrastructure, the lesson should be reinforcing: purpose-built hardware for predictable, high-volume workloads is the correct long-term architecture.

The tactical signal is the fourth. Miners with existing data center infrastructure, power relationships, and cooling capacity are positioned to participate in the AI infrastructure buildout as this market validates. Whether through mullet mining hybrid operations, direct GPU and ASIC colocation for neoclouds and enterprise AI customers, or full transition to AI compute infrastructure, the optionality is more valuable than it appears at first glance. The AI chip sector's $100 billion-plus annual CapEx has to land somewhere physical. Bitcoin mining operators own a lot of the physical infrastructure it needs.

The Inference Economy Has Many Winners

NVIDIA's $20 billion Groq acquisition was not the end of the independent AI chip story. It was confirmation that the inference economy is big enough to support multiple winners with different architectural approaches. Broadcom, targeting $100 billion in FY2027 AI revenue. Cerebras, targeting a $35 billion IPO in May 2026. Etched, valued at $5 billion on five major investors' conviction. Tenstorrent, at $3.2 billion on Jim Keller's bet that open ecosystems beat closed ones. Groq, proving the $20 billion acquisition premium is the new benchmark. Each thesis has real tensions. Each faces real execution risk. But the collective reality is that inference compute at scale is too large a market for NVIDIA to own alone. The market is pricing that reality aggressively.

The operational consequence is that the heterogeneous infrastructure standard (where companies maintain codebases across GPUs, ASICs, and LPUs simultaneously) is now the permanent architecture of production AI. Hyperscalers are already running it. Enterprise AI deployments will follow over the next eighteen months. The days of "pick one chip and build everything on it" are over. The days of multi-chip, multi-architecture, workload-optimized inference platforms are here.

The structural consequence is that the "custom silicon at scale" thesis has moved from speculative to validated at the largest technology companies in the world. Broadcom's financials alone validate the thesis at a scale that no pre-2024 forecast projected. That validation carries forward to every adjacent market where the same industrial logic applies, Bitcoin mining included. The custom silicon playbook works. The only open question is which specific workloads justify it.

Open Questions

The AI ASIC market has so many open questions.

The independent AI chip trade is not a moment. It's a multi-year structural transition. And it's still early.