The Second Silicon Line — 72 Hours Google Doubled the TPU Lead, Samsung Adopted OpenAI, and Qualcomm Anchored Amazon's $60B
A $760B depreciation clock is pushing hyperscalers off the merchant GPU and onto their own silicon — NVIDIA's moat is still wide, but the second line has now been drawn
Three custom-silicon anchors, one window
Over 72 hours the US semiconductor news quietly aligned around a single pattern: the hyperscaler decision to step outside the merchant GPU surfaced in three explicit announcements at once.
First, Google Cloud went on record that its own AI chip business — the TPU line — is twice the size of any rival hyperscaler's custom silicon. This is not a vanity metric. It is a P&L signal that a meaningful share of Google workloads is now earning revenue on Google's own silicon, not on NVIDIA GPUs.
Second, Samsung Electronics was reported to be partnering with OpenAI on a next-generation custom chip. Terms are undisclosed, but the direction is what matters: OpenAI, until now NVIDIA's largest single customer, is quietly digging its own silicon lane through the Samsung foundry.
Third, Qualcomm and Amazon are pursuing a CPU-centered AI alliance sized up to $60 billion. Amazon already runs Trainium and Inferentia; the Qualcomm tie adds an Arm server-CPU axis on top of that stack.
None of these three replaces NVIDIA GPUs. But all three route around them.
The clock pressing from above — $760B in depreciation
Inside the same three-day window, a countervailing story landed: AI chip depreciation is threatening the return on big tech's $760B AI capex program. Accelerators may sit on a 5-6 year accounting life, but from a workload perspective a generation turns in 2-3 years. On that depreciation curve, if per-GPU workload cost does not keep falling, the IRR that justified the capex breaks.
Custom silicon is the exact mechanism for rewriting that cost equation. Google's TPU business isn't twice its rivals because Google is showing off — it's twice its rivals because token-cost on TPU sits meaningfully below the merchant-GPU alternative on Google's own workloads. The same arithmetic is what lets Amazon commit $60B — not $6B — to Qualcomm.
SK Hynix reportedly considering a KRW 40 trillion (~$30B) share buyback in Q4 is the symmetric signal. When a memory supplier near cycle-peak is redirecting peak-cycle cash to shareholder returns rather than reinvestment, that supplier is quietly hedging the next-cycle reinvestment risk. Project Braid — the Google-Blackstone data-center JV — reportedly slipping on some builds points the same way.
NVIDIA's counter — widen the platform surface
NVIDIA is not idle. Inside the same 72 hours it played three counter-cards.
First, CUDA 13.4 shipped with Windows-on-Arm and next-gen Vera Rubin support. This physically extends CUDA lock-in past x86 into the Arm camp. The moment Amazon and Qualcomm push toward Arm CPUs, NVIDIA ensures CUDA still runs on top.
Second, NVIDIA announced a partnership with Australian cloud and data-center operators to stand up 2GW of AI infrastructure by 2027. India's Yotta locked in Vera Rubin at Noida as part of a $12B expansion. While US and Asian hyperscalers dig custom-silicon trenches, NVIDIA is directly claiming tier-2 sovereign AI capacity.
Third, at the Goldman Sachs tech conference NVIDIA re-anchored the Vera Rubin mass-production timeline and export-control posture — the market's re-verification point on generational volume.
Memory keeps running hot, which pushes custom silicon harder
DDR5 16Gb spot printed $54.33 on 2026-09-10. Samsung and SK Hynix memory inventory has reportedly crashed below 10 days. Apple raised iPhone pricing across the lineup on DRAM cost. Chinese AI chipmakers are hiking prices as HBM tightens. Intel is raising CPU prices 10% starting October.
With memory, CPU and HBM all rising in the same window, the BOM under an NVIDIA rack is being pushed up. That cost increment flows straight into hyperscaler workload cost, and when workload cost meets the depreciation curve the hyperscaler has to pick one of two paths: cut capex, or lift the custom-silicon share of the mix. Project Braid's delay looks like the first path. The Google-Samsung-OpenAI-Amazon-Qualcomm combination looks like the second.
Positioning
The long thesis on NVIDIA is intact — CUDA, Vera Rubin and sovereign AI keep pushing total volume higher. What this three-day window newly draws is the first quantified news-flow evidence that custom silicon's share of hyperscaler AI workload mix could climb into meaningful double-digit percent by 2028.
The second silicon line is not the line that replaces NVIDIA. It is the line that caps NVIDIA's growth rate.
Key Sources: - Google Cloud Says Its AI Chip Business Is Twice the Size of Any Rival Hyperscaler (BigGo, 2026-09-09) - Samsung working with OpenAI on next-generation custom chip (DataCenterDynamics, 2026-09-09) - Qualcomm and Amazon pursue $60B AI alliance centered on CPU development (MoneyNeverSleeps, 2026-09-09) - AI Chip Depreciation Threatens Returns on Big Tech's $760 Billion Capex Splurge (BigGo, 2026-09-09) - Nvidia partners with Australian companies to bring 2GW online by 2027 (DataCenterDynamics, 2026-09-10) - plus 12 more
If this analysis was helpful · ☕ Support Us · ✈️ Telegram