Will custom AI ASICs account for more than 20% of hyperscaler AI compute capacity by end of CY2026?

The Q4 earnings update materially shifts the ASIC landscape against the 20% threshold. Three developments compound: (1) Groq's $13B licensing deal eliminates the most prominent independent inference ASIC company — Jensen compared it to Mellanox, signaling strategic importance. Groq's low-latency inference tech now strengthens NVIDIA's stack in the exact workload domain where ASICs had the strongest competitive case. (2) Anthropic's $10B NVIDIA investment creates genuine dual-infrastructure, diluting what was the strongest TPU displacement signal (1M Google TPUs). Anthropic will now train and inference on Grace Blackwell and Vera Rubin, splitting workloads rather than consolidating on ASICs. (3) NVLink Fusion architecturally co-opts custom silicon — ASICs connected via NVLink become NVIDIA ecosystem components, not competitors. Even if hyperscalers deploy more custom chips, the measurement of 'ASIC compute independent of NVIDIA' becomes blurred. Against this, structural momentum continues: Amazon Trainium2 training frontier models, Google TPU v7 deployed, ASIC shipments still projected +44.6% growth. But reaching 20% of total compute by end CY2026 requires current deployed ASIC share to already be 12-15%, and NVIDIA's $62.3B Q4 DC revenue (+22% QoQ) is rapidly expanding the denominator. The math increasingly does not work within the timeframe.

I weight the structural ASIC momentum slightly higher than the consensus. Amazon Trainium2 training Claude 4 was the single most important piece of evidence in the original context — it proved frontier training is possible without NVIDIA GPUs. This fact has not changed. Google internally deploys TPU v7 for Gemini with claimed 40-60% lower TCO. Meta MTIA continues development. The Q4 data shows NVIDIA winning the investment and partnership battle, but it does not show ASIC programs being canceled. Anthropic's $10B commitment to NVIDIA does not eliminate its 1M TPU commitment to Google — it adds capacity. The question is whether these parallel deployments aggregate to 20% of total compute by CY2026 end. My concern is measurement uncertainty: no authoritative public source tracks this metric, and different methodologies could yield very different numbers. If someone counts Google's internal TPU deployment for all of Gemini's training and inference, plus Amazon's Trainium fleet for Claude and internal workloads, plus Meta's MTIA inference chips, the ASIC share could be higher than the market perceives based on NVIDIA's revenue dominance. Revenue share overstates compute share because NVIDIA GPUs are more expensive per unit of compute than some ASICs. Still, 20% by year-end is aggressive given NVIDIA's massive volume growth compressing the ratio.

The cumulative strategic moves in Q4 represent NVIDIA executing a deliberate ASIC containment strategy: acquire competing inference tech (Groq), invest in major ASIC customers to maintain GPU dependency (Anthropic $10B), and architecturally co-opt custom silicon (NVLink Fusion). This is not passive defense — it is active suppression of ASIC market share growth. The $17.5B in non-marketable equity purchases in FY2026 (up from $1.5B in FY2025) suggests this strategy extends well beyond Anthropic and Groq. NVIDIA is systematically buying positions in AI companies that might otherwise migrate to ASICs. For the 20% threshold: the original context estimated ASIC share is 'well below 20%' based on NVIDIA's 86%+ GPU market share. Q4 data shows NVIDIA gained share if anything — DC revenue grew +22% QoQ and +93% YoY. The denominator is expanding faster than any plausible ASIC numerator growth rate within the remaining ~10 months. Even with the +44.6% projected ASIC shipment growth, shipments must be deployed and operational to count as 'compute capacity.' The deployment lag means chips shipped in H2 2026 may not be fully operational by the snapshot date.

The new data systematically undermines the ASIC displacement narrative within the CY2026 timeframe. Groq's elimination as an independent competitor is significant — it was the highest-profile pure-play inference ASIC company. The Anthropic investment creates a dual-supplier dynamic that dilutes TPU concentration. NVLink Fusion is perhaps the most strategically important development: by enabling ASIC integration via NVIDIA's networking fabric, it transforms the competitive framing from 'ASIC vs GPU' to 'ASIC within NVIDIA ecosystem.' This architectural move could mean that even as ASIC silicon proliferates, NVIDIA captures the networking and software revenue. For the 20% question specifically: NVIDIA's Q4 DC revenue of $62.3B implies the company is deploying compute capacity at an extraordinary rate. The denominator of 'total hyperscaler AI compute capacity' is growing by billions of dollars per quarter. For ASICs to reach 20% of this expanding total, ASIC deployment would need to be roughly $15-16B worth of compute equivalent — a figure that strains credulity for CY2026. Meta committed to 'millions of Blackwells and Rubins,' OpenAI deploying GPT-5.2-Codex on GB NVL72 — the major AI companies are deepening GPU commitments, not shifting away.

While the Q4 data clearly shifts the balance toward NVIDIA in the short term, I want to flag that the structural ASIC adoption trend has not reversed — it has been partially co-opted and partially delayed. Google TPU v7 is in production and scaling. Amazon Trainium2 has proven frontier training capability. These are not speculative programs; they are deployed infrastructure. The question is whether their aggregate share reaches 20% by CY2026 end. The prior context estimated current deployed share as 'well below 20%' — likely in the 8-12% range based on NVIDIA's GPU dominance but accounting for Google's substantial internal TPU fleet. If current share is closer to 12%, the +44.6% shipment growth rate could push deployed capacity toward 15-17% by year-end. If current share is closer to 8%, reaching 20% requires a step-function increase that does not appear in the data. The Groq and Anthropic deals reduce the probability but do not eliminate it. Google's TPU deployment for Gemini is the single largest ASIC compute base and was not affected by Q4 developments. Jensen's notably minimal ASIC discussion in Q4 could signal either confidence or deliberate de-emphasis — I weight it as neutral. Overall, I place probability slightly above the low end of the revised range given measurement uncertainty.

The convergence of strategic moves in Q4 is decisive for the near-term ASIC share question. Consider the math: NVIDIA guided Q1 FY2027 at $78B total, implying ~$68-70B DC revenue. If NVIDIA maintains this trajectory through CY2026, annual DC revenue would be approximately $280-300B. For ASICs to represent 20% of total compute capacity, ASIC compute equivalent would need to be ~$70-75B — roughly 25% of the remaining compute pie. While Google, Amazon, and Meta all deploy custom silicon, their combined ASIC spending is likely in the $20-30B range for CY2026, representing perhaps 7-10% of total compute capacity rather than 20%. The NVLink Fusion architecture further complicates this: if AWS deploys Trainium chips connected via NVLink, does that count as ASIC compute or NVIDIA ecosystem compute? This definitional ambiguity makes the 20% threshold harder to breach in practice. The Groq acquisition is particularly impactful because it removes the most credible independent inference optimization path — inference is the workload growing fastest and where ASICs had the strongest TCO argument.

Q4 FY2026 data provides three clear signals against ASICs reaching 20% by CY2026 end: Groq acquired (independent competitor eliminated), Anthropic locked into dual GPU/TPU infrastructure ($10B investment), and NVLink Fusion co-opts custom silicon. NVIDIA's DC revenue at $62.3B with +22% QoQ growth massively expands the denominator. The 20% threshold requires ASIC compute to grow faster than NVIDIA's GPU compute — and NVIDIA is accelerating, not decelerating. Prior 20% probability was already at the low-moderate end; new data shifts it lower.

The ASIC 20% threshold is a near-term question (CY2026 end, ~10 months away) against a backdrop of NVIDIA's strongest-ever quarterly performance. Even with projected +44.6% ASIC shipment growth, the absolute deployment volume cannot plausibly reach 20% of total hyperscaler AI compute when NVIDIA is shipping at the current rate. Google's internal TPU deployment is the largest ASIC base, but Google also purchases substantial NVIDIA GPU capacity. Amazon Trainium2 is deployed but at much smaller scale than AWS's GPU fleet. Meta MTIA is still early. The combined ASIC deployment across all hyperscalers is likely 10-12% of total compute capacity currently, and reaching 20% in 10 months requires a doubling that contradicts the deployment timelines for new ASIC capacity.

The structural ASIC trend is real but the 20% by CY2026 end timeline is aggressive. Q4 data reinforces this: NVIDIA's massive volume growth, the Groq acquisition, Anthropic dual-infrastructure, and NVLink Fusion all work against ASIC share growth in the near term. The measurement uncertainty is the wild card — if industry research defines 'AI compute capacity' narrowly (e.g., inference-only at certain hyperscalers), ASIC share could be higher than the headline GPU revenue dominance suggests. But the consensus measurement approach would count total deployed compute capacity across training and inference, where NVIDIA's dominance is most pronounced. Probability is in the 13-17% range, I lean toward the middle given measurement uncertainty.