The Dual-Track TPUv9 Program: Humufish and Triggerfish

Google's TPUv9 generation is now a two-pronged program :

• Humufish: The base TPUv9 platform, designed and supplied by Broadcom. JPMorgan and other analysts confirm this program remains on track for volume production in calendar 2028 using a 2-nanometer process at TSMC .
• Triggerfish: An upgraded, higher-priced variant built on the Humufish platform but exclusively designed by MediaTek . It is positioned not as a replacement, but as an incremental addition targeting specialized, latency-sensitive AI workloads .

This dual-track approach lets Google optimize for different workload profiles while mitigating single-vendor dependency.

Architectural Upgrades: A Unified Chip for Agentic AI

The Triggerfish chip represents a significant architectural departure from the TPUv8 family :

• Unified Training and Inference: Unlike previous generations that separated training and inference accelerators, Triggerfish combines both capabilities into a single package. It is designed specifically for agentic AI and reinforcement learning workloads .
• 2–3x Larger SRAM: The on-chip SRAM cache is significantly expanded, allowing more active working sets to be kept locally on the TPU. This reduces data movement and improves efficiency during low-latency decoding .
• HBM4E Memory: Triggerfish upgrades off-chip memory from HBM4 (used in the Humufish base design) to the next-generation HBM4E standard, providing even higher bandwidth .
• New CPU Tile (Simulation Die): A dedicated CPU tile is integrated into the same package as the main compute die. Sometimes called a "simulation die," this component is designed to handle local TPU management, training/inference mode switching, and the coordination of multiple AI agents .

These upgrades are aimed at breaking through the "CPU wall" and "memory wall" that constrain AI workload performance .

Production Timeline and Volume

According to Kuo's supply-chain survey, the timeline for Triggerfish is as follows :

• Production start: Late 2027 (Q4 2027)
• Volume ramp: 2028
• Lifecycle volume: 1–2 million units, incremental to the 4–5 million units expected for the base Humufish design .
• Unit price: Approximately 30% higher than the base Humufish variant .

The chip is being fabricated through MediaTek's partnership with TSMC .

Broadcom's Secured Role: A Long-Term Agreement Through 2031

Despite MediaTek's win on Triggerfish, Broadcom's relationship with Google is far from over. An SEC 8-K filing from June 29, 2026, formally discloses a Long Term Agreement for Broadcom to develop and supply custom TPUs for Google's future generations, plus a Supply Assurance Agreement for networking components, through 2031 . Analysts confirm this covers four TPU generations .

JPMorgan and Morgan Stanley have explicitly refuted market rumors of any delay or cancellation of Broadcom's TPUv9 program, stating it "remains on track with no delays" . Broadcom's v9 variant, also targeting 2028 volume production on a 2nm TSMC process, handles the base Humufish design . Some analysts note Broadcom's variant may face disadvantages in HBM capability and process-node timing relative to the MediaTek-enhanced design .

Multi-Supplier Diversification as a Strategic Imperative

Google's TPUv9 strategy reflects a broader industry trend toward multi-supplier diversification. By splitting the base and enhanced variants between Broadcom and MediaTek, Google reduces its reliance on any single vendor while optimizing for different cost and performance points . This also creates competitive tension that can drive innovation and pricing efficiency.

Impact on Broadcom: Strong AI Growth, but High Expectations Punished

The market's reaction to Broadcom's fiscal Q2 2026 earnings on June 3, 2026, illustrates the intense scrutiny on AI chip suppliers :

• Q2 Revenue: $22.19 billion, missing the $22.27 billion consensus .
• AI Chip Revenue: $10.8 billion (up 143% YoY) .
• Q3 AI Chip Revenue Guidance: $16 billion, missing the sell-side consensus of $17.2 billion by ~7% .
• Stock Reaction: Shares fell 13–15% in extended trading, wiping out approximately $50 billion in market cap .

Key Nuance: Broadcom's $16 billion Q3 AI guidance still represented over 200% year-over-year growth . The selloff reflected that the market had priced AI semiconductor stocks "for perfection," punishing even strong double-digit growth that fell below elevated expectations . CEO Hock Tan declined to raise the previous FY27 AI revenue target of ~$100 billion, which further disappointed bullish bets .

Summary

Google's custom AI chip strategy is evolving from a single-vendor model to a sophisticated dual-track approach. MediaTek's exclusive contract for the enhanced Triggerfish variant represents a major win for the Taiwanese semiconductor firm, while Broadcom remains the foundational long-term partner for Google's TPU roadmap through 2031. The architectural innovations in Triggerfish—unified training/inference, dramatically larger SRAM, HBM4E memory, and a co-packaged CPU tile—directly target the demands of agentic AI. For investors, the Broadcom earnings reaction serves as a reminder that even dominant AI chip suppliers face volatile market expectations.

The Dual-Track TPUv9 Program: Humufish and Triggerfish

Google's TPUv9 generation is now a two-pronged program :

• Humufish: The base TPUv9 platform, designed and supplied by Broadcom. JPMorgan and other analysts confirm this program remains on track for volume production in calendar 2028 using a 2-nanometer process at TSMC .
• Triggerfish: An upgraded, higher-priced variant built on the Humufish platform but exclusively designed by MediaTek . It is positioned not as a replacement, but as an incremental addition targeting specialized, latency-sensitive AI workloads .

This dual-track approach lets Google optimize for different workload profiles while mitigating single-vendor dependency.

Architectural Upgrades: A Unified Chip for Agentic AI

The Triggerfish chip represents a significant architectural departure from the TPUv8 family :

• Unified Training and Inference: Unlike previous generations that separated training and inference accelerators, Triggerfish combines both capabilities into a single package. It is designed specifically for agentic AI and reinforcement learning workloads .
• 2–3x Larger SRAM: The on-chip SRAM cache is significantly expanded, allowing more active working sets to be kept locally on the TPU. This reduces data movement and improves efficiency during low-latency decoding .
• HBM4E Memory: Triggerfish upgrades off-chip memory from HBM4 (used in the Humufish base design) to the next-generation HBM4E standard, providing even higher bandwidth .
• New CPU Tile (Simulation Die): A dedicated CPU tile is integrated into the same package as the main compute die. Sometimes called a "simulation die," this component is designed to handle local TPU management, training/inference mode switching, and the coordination of multiple AI agents .

These upgrades are aimed at breaking through the "CPU wall" and "memory wall" that constrain AI workload performance .

Production Timeline and Volume

According to Kuo's supply-chain survey, the timeline for Triggerfish is as follows :

• Production start: Late 2027 (Q4 2027)
• Volume ramp: 2028
• Lifecycle volume: 1–2 million units, incremental to the 4–5 million units expected for the base Humufish design .
• Unit price: Approximately 30% higher than the base Humufish variant .

The chip is being fabricated through MediaTek's partnership with TSMC .

Broadcom's Secured Role: A Long-Term Agreement Through 2031

Despite MediaTek's win on Triggerfish, Broadcom's relationship with Google is far from over. An SEC 8-K filing from June 29, 2026, formally discloses a Long Term Agreement for Broadcom to develop and supply custom TPUs for Google's future generations, plus a Supply Assurance Agreement for networking components, through 2031 . Analysts confirm this covers four TPU generations .

JPMorgan and Morgan Stanley have explicitly refuted market rumors of any delay or cancellation of Broadcom's TPUv9 program, stating it "remains on track with no delays" . Broadcom's v9 variant, also targeting 2028 volume production on a 2nm TSMC process, handles the base Humufish design . Some analysts note Broadcom's variant may face disadvantages in HBM capability and process-node timing relative to the MediaTek-enhanced design .

Multi-Supplier Diversification as a Strategic Imperative

Google's TPUv9 strategy reflects a broader industry trend toward multi-supplier diversification. By splitting the base and enhanced variants between Broadcom and MediaTek, Google reduces its reliance on any single vendor while optimizing for different cost and performance points . This also creates competitive tension that can drive innovation and pricing efficiency.

Impact on Broadcom: Strong AI Growth, but High Expectations Punished

The market's reaction to Broadcom's fiscal Q2 2026 earnings on June 3, 2026, illustrates the intense scrutiny on AI chip suppliers :

• Q2 Revenue: $22.19 billion, missing the $22.27 billion consensus .
• AI Chip Revenue: $10.8 billion (up 143% YoY) .
• Q3 AI Chip Revenue Guidance: $16 billion, missing the sell-side consensus of $17.2 billion by ~7% .
• Stock Reaction: Shares fell 13–15% in extended trading, wiping out approximately $50 billion in market cap .

Key Nuance: Broadcom's $16 billion Q3 AI guidance still represented over 200% year-over-year growth . The selloff reflected that the market had priced AI semiconductor stocks "for perfection," punishing even strong double-digit growth that fell below elevated expectations . CEO Hock Tan declined to raise the previous FY27 AI revenue target of ~$100 billion, which further disappointed bullish bets .

Summary

Google's custom AI chip strategy is evolving from a single-vendor model to a sophisticated dual-track approach. MediaTek's exclusive contract for the enhanced Triggerfish variant represents a major win for the Taiwanese semiconductor firm, while Broadcom remains the foundational long-term partner for Google's TPU roadmap through 2031. The architectural innovations in Triggerfish—unified training/inference, dramatically larger SRAM, HBM4E memory, and a co-packaged CPU tile—directly target the demands of agentic AI. For investors, the Broadcom earnings reaction serves as a reminder that even dominant AI chip suppliers face volatile market expectations.