The USD 725B Question: As the Big Four CSPs Push 2026 Capex to Astronomical Levels, Optical Communications' Real Bottleneck Quietly Moves Upstream (W21)
This week NVIDIA, Cisco, and the Big Four CSPs together pushed the AI compute ceiling up another notch. But open up the Chinese module makers' earnings calls and a "revenue up, profit lagging" signal tells you that what's truly scarce in this bull market is not at the module assembly stage.
Starting with an Easily Overlooked Contrast
Putting two of this week's numbers side by side is more direct than any analysis.
The first is NVIDIA: FY27 Q1 revenue of USD 81.6B, Data Center at USD 75.2B, and Q2 guidance of USD 91B. Jensen Huang declared on the call that "agentic AI has arrived and demand is parabolic" (CNBC). It also raised its buyback to USD 80B and lifted the quarterly dividend from USD 0.01 to USD 0.25 in one go—25x. A company answering the question "will we be short of cash?" with a shareholder return program.
The second is Eoptolink: Q1 2026 revenue of RMB 8.338B, YoY +106%; but its full-year net profit growth (+236% for 2025) has clearly outpaced revenue (+60%) (STCN). Innolight was even more extreme, with Q1 revenue of RMB 19.496B, YoY +192%. The combined market cap of the three leading Chinese optical module makers has passed RMB 1 trillion.
The first set of numbers tells you AI compute demand hasn't peaked; the second should belong to the biggest beneficiaries. But spread out the gross margins and an interesting tension appears: while downstream module makers' revenue is still doubling, margin expansion can no longer keep pace.
This weekly report aims to answer two questions: first, with NVIDIA, Cisco, and Big Four CSP capex all revised up this week, where has the industry been taken? Second, while the demand ceiling keeps rising, the real supply bottleneck has quietly shifted from "module capacity" upstream—what does this mean for Taiwanese suppliers?
Thread One: AI Demand Is Finally Spreading from Hyperscalers to Enterprises
Starting Point: The Big Four CSPs Lift 2026 Capex to USD 725B
For two years, hyperscaler capex has been at the center of the debate over "should the market worry about a 2H soft landing." This week that debate temporarily lost its footing.
Combined CY26 capex of the Big Four CSPs has been raised to USD 725B, YoY +77% (Tom's Hardware). Broken down: Microsoft USD 190B, Google USD 180–190B, Meta USD 125–145B, AWS USD 200B. Microsoft's CFO specifically noted that USD 25B of this was forced by rising memory and chip costs—a remark worth amplifying. It means part of the capex increase is "passive": not optimism, but supply so tight that more money is needed to lock it in.
Even more notable is a structural shift: more than 60% of spending is going to power, cooling, and data center construction, not compute itself. This trend has become clearer in every hyperscaler call over the past three quarters, but this week's numbers pushed it to a share that can't be ignored. For STT readers, it means capex beneficiaries are broadening from pure compute (GPU / ASIC) to power (Delta Electronics), thermal (Auras Technology), and high-speed interconnect (optical / cable).
Evidence: Cisco Doubles Its AI Order Outlook
If the capex numbers can still be read as "the hyperscalers' own script," Cisco's call is a more industry-significant inflection signal.
Cisco's Q3 FY26 revenue was USD 15.8B, YoY +12%—not earth-shattering on its own. But AI infrastructure orders alone reached USD 1.9B in the quarter, and the full-year AI order outlook jumped from USD 5B to USD 9B—doubling, and four times all of FY25 (Futuriom). Silicon One shipments passed one million units, and the G300 (102.4Tbps) is starting to win design-ins.
CEO Chuck Robbins said something after the call that spelled out the industry's pecking order bluntly: "Without silicon, you'll become irrelevant to hyperscalers (Benzinga)." On the surface it's marketing talk; in substance it's a direct shot at Arista and Juniper.
But what deserves more discussion is the "source" of Cisco's order growth. Cisco is not hyperscalers' first-choice networking vendor—its core customers are enterprises, second-tier clouds (neoclouds), and sovereign AI. When its AI orders hit USD 1.9B in a single quarter and the full-year outlook doubles, the signal is: AI compute demand is finally starting to spread from hyperscalers to enterprises and neoclouds. For the past 24 months hyperscalers were the sole engine of AI capex; now a second engine is starting up.
Interpretation
Looking only at NVIDIA's call, the story is simple—AI compute demand remains strong. But stacking NVIDIA, Cisco, and Big Four CSP capex together gives a richer picture: the structure of compute demand is shifting from "training clusters concentrated at a few hyperscalers" to "multi-scenario deployments spreading into inference, enterprises, and neoclouds."
This has two direct implications for the optical communications and AI infrastructure supply chain. First, unit demand for optical modules and high-speed interconnect will be larger than estimates based on hyperscalers alone—enterprise and neocloud deployments involve more racks at smaller individual scale, so total port counts grow non-linearly. Second, the 800G-to-1.6T generational transition may move faster than module makers' original capacity plans, and supply tightness in the second half will be worse than now.
This thread leads us to the second story.
Thread Two: The Real Bottleneck Has Moved from Module Assembly Upstream
Starting Point: AOI Targets 700K Units per Month of 800G Capacity by 2027
Applied Optoelectronics (AOI)'s Q1 call was the most easily overlooked this week, yet the most directly relevant to Taiwanese suppliers.
In Q1 2026, AOI completed its first volume 800G shipments to a large hyperscaler and won USD 53M in new orders, with Q1 800G monthly capacity of nearly 100K units (SEC 8-K). Its Houston footprint has nearly doubled, targeting 800G+1.6T monthly capacity of 700K units by end-2027 and another +350% in laser wafer capacity. AOI itself expects 800G/1.6T demand to exceed capacity at least through mid-2027.
Put this in context: AOI's turnaround from years as "the Houston alternative to Chinese module makers" into a tier-one hyperscaler supplier is the most representative case of US optical onshoring in recent years. But what's more worth tracking is its upstream—the 350% expansion of Houston laser wafer capacity drives PD/LD orders for GCS Holdings. AOI and Amazon have signed a 10-year USD 4B contract, and GCS's position in this supply chain is nearly irreplaceable.
Evidence: A Structural "Revenue Up, Profit Lagging" Signal Appears in Chinese Optical Modules
The Big Four Chinese optical module makers all reported stunning revenue growth this week. Innolight's Q1 revenue was RMB 19.496B, YoY +192%; Eoptolink's Q1 revenue was RMB 8.338B, YoY +106% (STCN).
But look closer and a structural problem emerges: as revenue doubles, net profit growth is being caught by revenue growth. When Eoptolink's 2025 net profit rose +236% on revenue +60%, gross margin was still expanding; in Q1 2026 the gap between profit and revenue growth is narrowing. In one sentence: module makers' margins are being eaten by upstream EML/InP price increases.
LandMark Optoelectronics validates this signal—customer demand for 800G is strong, 1.6T has been shipping since 2H 2025, and gross margin has surged to 48%. NVIDIA's direct NTD 40B stake is not a strategic investment but a "capacity lock" signal. When downstream module makers' margins are being squeezed while upstream InP component makers' margins expand, it's clear where the spread is going.
One more: VPEC's Q1 2026 revenue was NTD 959M, YoY +20.87%, with gross margin of 38.09%—strong even in the off-season (TechNews). VPEC's pivot from 5G PA to optical epitaxy will only fully play out when 1.6T truly ramps in 2H 2026.
Interpretation: When Three Signals Intersect
Module makers' "revenue up, profit lagging," AOI's orders full through mid-2027, LandMark's margin hitting 48%—once these three signals intersect, the real order of beneficiaries in this optical bull market becomes clear:
The biggest beneficiaries are upstream EML / InP / PD / LD and epitaxy (LandMark, GCS Holdings, VPEC); second come OSAT and opto-electronic co-testing (ASE). Module makers themselves will be caught between "expanding capacity to catch demand" and "upstream price increases."
The practical takeaway for readers: if you've focused only on module makers for the past 12 months, it may be time to reallocate research time to upstream components. The optical industry's "smile curve" will be steeper in the 1.6T generation than in 800G—value is growing at both ends, while assembly in the middle is squeezed.
Other Signals Worth Tracking This Week
Intel (INTC): CEO Lip-Bu Tan publicly said 18A yields are improving 7-8% per month, multiple external customers are in talks, 14A is positioned against TSMC's equivalent node, and hinted at new collaboration items with NVIDIA (CNBC). If 18A truly secures external commitments in the second half, it would be the biggest shift in the foundry landscape in five years; TSMC's pricing power will only start being tested in the second half.
TSMC (2330): The 5.5-reticle CoWoS platform enters volume production this year with yields >98%; CoWoS capacity will rise from 35K wpm at end-2024 to 120–140K wpm by end-2026, roughly 4x (DigiTimes). The CoWoS ramp paves the way for NVIDIA Rubin / Vera Rubin; but HBM is the real key variable for second-half supply bottlenecks.
Astera Labs (ALAB): Q1 2026 revenue of USD 308.4M, YoY +93%; the Scorpio X-Series 320-lane fabric switch has begun shipping, with Q2 guidance of USD 355–365M (Astera Labs IR). Scorpio X marks ALAB's leap from "signal conditioning supplier" to "scale-up fabric architect," positioned against NVLink Switch + UALink.
NVIDIA CPO platform timeline confirmed: Quantum-X InfiniBand CPO launches in 1H 2026 (115 Tb/s switch, 144 ports × 800G), and Spectrum-X Photonics (Ethernet CPO) in 2H 2026 (NVIDIA Developer Blog). CPO does not "replace" pluggables but runs "in parallel"—XPU-side CPO is expected only in 2027, so 1.6T pluggable module makers still have at least two good years.
GlobalFoundries (GFS): In May it unveiled SCALE (Silicon Photonics Co-packaged Advanced Light Engine), an OCI MSA platform; its silicon photonics business is expected to double to USD 400M in 2026 and surpass USD 1B by 2028 (TrendForce). OCI MSA is a fabric-level extension of the UALink camp, forming a second ecosystem alongside NVIDIA Quantum-X / Spectrum-X.
STT Observations and What to Watch Next Week
This week's takeaway fits in one sentence: as the demand ceiling is pushed to USD 725B, what's truly scarce on the supply side has shifted from module capacity to upstream components. The NVIDIA and Cisco calls tell you demand hasn't peaked and enterprises are taking the baton; the calls from the four Chinese module makers, AOI, and LandMark tell you where the money will flow. Stacked together, these two threads give readers a more three-dimensional supply chain map than any single company could.
Three things deserve real attention next.
Whether Intel announces substantive external customers between COMPUTEX Taipei in June and SEMICON West in the US in July—if so, TSMC's pricing power in the 2nm/A14 generation will face its first serious challenge.
The timing of Delta Electronics' 800V HVDC platform alignment with NVIDIA's Rubin platform, and whether small-volume shipments in 2H 2026 arrive on schedule.
Under the dual CPO ecosystems (NVIDIA's own vs. UALink + GFS SCALE), the positioning order of FOCI, LandMark, and TSMC COUPE—next quarter's earnings calls will be the key watch point.
The optical communications story will not be linear in the 1.6T generation. As demand structure spreads from training to inference, from hyperscalers to enterprises, and from compute to power and interconnect, research time allocation should change accordingly.
*STT Market Insights | 2026-05-25 | Compiled by the STT editorial team*
Comments