The Same Week NVIDIA Earned $96.2 Billion, It Told Customers Prices Are Going Up 15%: Optical Communications' Bottleneck Has Shifted from Capacity to BOM — 2026 W35 Industry Weekly (08/24–08/31)
NVIDIA FY27 Q2 revenue was $96.2 billion, with data center at $89.0 billion, YoY +117%, and next-quarter guidance of $108 billion; Marvell posted a record $2.739 billion for the quarter, guiding QoQ +15% for next quarter. There was no doubt on the demand side this week.
That same week, NVIDIA notified Microsoft, Google, and Oracle that AI servers shipping in early 2027 will cost more than 15% more, driven by memory, not GPUs.
Coherent prepaid AXT $22.3 million in exchange for three years of committed 6-inch InP capacity — the substrate shortage has already forced companies to "pay cash to lock in capacity."
China's "Yi-Zhong-Tian" trio (Eoptolink, Innolight, TFC Communication) earned a combined net profit of about RMB 22.3 billion in the half year, but because the FCC draft "exempts models already sold in the US," near-term order shifts will be far smaller than the market imagines.
1. Why This Week: Demand and Cost Both Hit Records Within 48 Hours
08/26 to 08/27 was the most concentrated earnings window of the year.
NVIDIA FY27 Q2 (ended 2026/7/26) revenue was $96.2 billion, YoY +106% and QoQ +18%; data center revenue was $89.0 billion, YoY +117%; gross margin was 75.0%, and non-GAAP EPS was $2.22. Next-quarter guidance is $108 billion, plus or minus 2%, above the market's expected $104.2 billion, and management also called for roughly 70% revenue growth in FY2028. The details matter more than the totals: Blackwell accounts for nearly 70% of data center compute revenue, meaning the Rubin-generation pull-in has not really started yet. (See Earnings Highlights: NVIDIA | Q2 FY2027)
Marvell followed the next day: FY27 Q2 revenue hit a record $2.739 billion, YoY +37%, with data center accelerating to +46%; next-quarter guidance of $3.15 billion implies QoQ +15%, the steepest among all AI semiconductor names this quarter. Management said optical interconnect revenue will grow more than 70% YoY in FY2027. (See Earnings Highlights: Marvell | Q2 FY2027)
Yet in the same week, news on another front looked nothing like good news: through its contract manufacturers, NVIDIA told major customers that Vera Rubin and Grace Blackwell systems shipping in early 2027 will cost more than 15% more, mainly because DRAM, HBM, and LPDDR costs are spiraling out of control. This comes on top of the price increase in July.
Record demand and runaway costs compressed into the same week is no coincidence.
2. This Week's Hard Numbers
Company | Figure | Notes |
NVIDIA | $96.2 billion | FY27 Q2 total revenue, YoY +106% |
NVIDIA | $89.0 billion | FY27 Q2 data center revenue, YoY +117% |
NVIDIA | $108 billion | FY27 Q3 revenue guidance (±2%) |
NVIDIA | More than 15% | Price increase on AI servers shipping in early 2027 |
Marvell | $2.739 billion | FY27 Q2 revenue (record), YoY +37% |
Marvell | $3.15 billion | FY27 Q3 revenue guidance (QoQ +15%) |
Marvell | More than 70% | FY2027 optical interconnect revenue YoY growth target |
Coherent | $22.3 million | Prepayment to AXT for three years of committed 6-inch InP capacity |
Lumentum | More than 30% | Amount by which AI laser demand exceeds supply |
Innolight | RMB 41.778 billion | 2026 H1 revenue, YoY +182.49% |
Eoptolink | RMB 20.91 billion | 2026 H1 revenue, YoY +100.34% |
TFC Communication | RMB 2.828 billion | 2026 H1 revenue, YoY +15.15% |
TSMC | 125–130k wafers/month | CoWoS capacity at end-2026 (YoY approx. +80%) |
TSMC | 15k wafers/month | Silicon photonics capacity in Q4 2026 |
AWS | 2 million units | Additional NVIDIA GPUs to be deployed in 2027–2028 |
Top 5 CSPs | $660–690 billion | Combined 2026 capex guidance |
3. Three Simultaneous Cost Shifts
Lay out this week's signals and three things point in the same direction:
Memory is competing with optics for the same BOM budget pool. The top five CSPs announced their 2026 capex guidance well before the price increase notice; with the ceiling unchanged and per-system cost up more than 15%, the number of racks they can buy will shrink, and optical modules are one of the easiest items in the system BOM to squeeze on price.
Pricing power over InP is consolidating at the very top of the supply chain. Coherent's willingness to prepay to lock in three years of substrate capacity means management judges this tightness will last at least until 2028. Lumentum is currently the only supplier shipping 200G/lane EMLs in volume, and says AI laser demand exceeds supply by more than 30%. For background, see You Can Ban the Modules, But Not the Substrates: Optical Communications' Decisive Battle Has Moved Down to Indium Phosphide.
400G/lane puts an expiration date on the EML shortage narrative. Once Broadcom's Taurus 400G/lane optical DSP lands, the laser count per 1.6T module will be cut in half.
As for China's "Yi-Zhong-Tian" RMB 22.3 billion, and the FCC draft's clause "exempting models already sold in the US" — the best earnings and the worst policy risk landed in the same week, and the market has quite possibly read the latter backwards.
──────────────────────────────
That concludes the key takeaways of this article.
STT's full interpretation — why rising memory prices will squeeze optics margins, what will validate the first reverse signal in InP long-term contracting, which quarter marks the expiration date for 400G/lane, and why the FCC's "existing models exemption" makes near-term order shifts far smaller than the market imagines — is fully analyzed in the premium section.
Subscribe to the STT premium section and read the full analysis on vocus (in Chinese):
Comments