CRDO — Customer Revenue Reconstruction

Date: 2026-03-03 | Analyst: wsm007 | Type: Supplementary Analysis

Customer Identity Map

Key sources:

• SEC: Amazon warrant in 10-K (crdo-20250503)
• Transcripts: Q3 FY25 (CFO named Amazon), Q2 FY26 (CFO: "#2 at 24% = first hyperscaler to ramp")
• Conferences: Barclays Sept 2025 (Microsoft named by analyst, xAI named by CEO), Goldman Sept 2025 (Meta confirmed)
• Analysts: Needham (Quinn Bolton) confirmed Amazon, Microsoft, xAI
• Media: CNBC "500 purple cables" (Oct 2025), Benzinga analyst roundup (Mar 2025)

Cross-Quarter Linkages (from Q3 FY26 transcript)

• Q3 Top 1 (39%) = "same customer as Q1's largest" + "quite a large increase Q2→Q3" → xAI
• Q3 Top 2 (32%) = "was largest customer in Q2" → Amazon
• Q3 Top 3 (17%) = "first hyperscaler to ramp" → Microsoft
• Q2 Top 2 (24%) = "first hyperscaler to ramp" (from Q2 transcript) → Microsoft

Per-Customer Revenue Reconstruction ($M)

Revenue by Customer

Revenue Share (%)

Key Insights

1. xAI Is the Wild Card — and the Biggest Growth Driver

$7M → $20M → $78M → $54M → 159M.MostvolatilecustomerrampinanycompanyIfollow.Colossusbuildoutcomesinwaves—massiveordersperdeploymentphase, thenpause, thensurge.Q2 → Q3jumpof+105M QoQ drove most of the $139M sequential revenue add. Brennan's "quite a large increase" was diplomatic.

2. Amazon Is Large but Steadier Than Position Rank Suggests

$81M → $104M → $74M → $80M → 130M.DippedQ1 − Q2(AWSdeploymentcycle), surgedQ3. 500M+ annual run-rate. $201M cumulative warrant fully vested — relationship deep and multi-year. Uses AECs for both back-end and front-end connections.

3. Microsoft Is the Foundation

$20M → $19M → $45M → $67M → $69M. Steady, predictable, upward. "First to ramp" customer provides baseline revenue. Share declining in % terms (15% → 17% of a much larger base) because others grow faster. In absolute dollars: 3.5x in four quarters. Front-end AEC adoption extending to back-end AI clusters.

4. Meta Is Early Innings

Crossed 10% briefly in Q2 (32M), heldflat(33M) in Q3. OCP Summit and Goldman conference confirm real engagement. Potential next xAI-type surge if Llama infrastructure buildout accelerates.

5. The Concentration Risk Is Really an xAI Risk

When top-3 = 88%, it's because xAI's orders are so massive and lumpy they periodically dominate the mix. If Colossus 2 build completes and orders pause, ~$159M could drop to $50-80M in a single quarter. That's the real risk embedded in the 88% number.

Bull vs Bear from This Table

Bull: Four confirmed hyperscalers ramping simultaneously, 5th starting. Amazon and Microsoft provide steady baseline. xAI lumpiness masks the fact that ALL customers are growing independently. If Meta follows xAI's trajectory with a 6-month lag, another $100M+ quarterly customer is emerging.

Bear: xAI concentration is extreme and unpredictable. One Colossus build pause = ~$100M quarterly revenue swing. "Diversification" narrative is partially illusion — three customers do 88% and the lumpiest one drives QoQ variance.

WSM verdict: Bull case stronger because all four customers are independently growing. xAI lumpiness creates volatility, not vulnerability. Infrastructure buildouts take years. But need Meta to cross $50M/quarter and 5th hyperscaler to reach 10% before concentration problem is solved.

Hyperscaler Infrastructure Buildout — What's Coming for Each Customer

xAI — Colossus Complex (Memphis/Southaven)

Funding: $20B Series E (Jan 2026), $10B round (Sep 2025), NVIDIA $2B equity stake, Tesla $2B stake. Stated 2026 infrastructure spend: "at least $30B."

CRDO revenue implications:

• Current wave (Q3 FY26 = $159M): Colossus 2 GPU installation driving massive AEC orders. 550K GB200s need rack-to-switch AEC connections. This is the wave we're in now.
• Next 2-3 quarters: Colossus 2 installation continues through Q3 2026 (CRDO's Q4 FY26 and Q1 FY27). Expect sustained $100-160M/quarter orders during active install.
• Building 3 wave: Conversion starts Q1 2026, GPU installation likely H2 2026/H1 2027. This creates a second wave of orders potentially overlapping with tail end of Colossus 2.
• Risk: Gap between Colossus 2 completion and Building 3 readiness could create a $50-80M dip quarter. But the building overlap suggests this gap may be minimal.
• Bull case: 1M GPU target means Building 3 adds ~220-320K more GPUs beyond current install. At ~100-150 AEC cables per NVL72 rack (72 GPUs), that's another ~300-670K cables. Multi-quarter sustained demand.

→ xAI demand runway: 4-6 more quarters of elevated orders (through mid-FY28), with possible brief dips between build phases.

Amazon (AWS) — The Multi-Gigawatt Machine

Key facilities:

• Project Rainier (Indiana): ~500K Trainium2, 1.3GW+, 30 buildings on 1,200 acres. Operational.
• Pennsylvania: $20B "AI innovation campuses"
• North Carolina: $10B build
• Louisiana: $12B campus
• Federal/Defense: $50B committed, ~1.3GW AI/HPC for government
• OpenAI: 2GW of Trainium3/4 under $50B contract

CRDO revenue implications:

• $200B CY2026 capex is nearly double 2025. Even if AECs are a fraction of that, the absolute dollars flowing to connectivity are massive.
• Capacity doubling (15GW → 31GW by end-2027) means the network infrastructure roughly doubles. Every new rack needs AEC connections.
• Trainium clusters use NVIDIA-style Ethernet networking (Spectrum-X equivalent), which means AECs apply to both NVIDIA and Trainium deployments.
• AWS is diversifying silicon but NOT diversifying networking architecture. AEC demand is agnostic to whether it's NVIDIA or Trainium inside the rack.
• Current $130M/quarter → expect $150-200M/quarter by FY27 as capacity ramp accelerates. Steady, not lumpy.

→ Amazon demand runway: Multi-year, steadily growing. The $200B capex year is just starting. Expect Amazon to be a $600-800M annual CRDO customer by FY28.

Microsoft — Power-Constrained but Doubling

Key dynamics:

• Power is the constraint, not GPUs. Nadella: "We have AI GPUs sitting in inventory because we lack the power to install them." This means orders to CRDO may be paced by power availability rather than GPU procurement cycles.
• Wisconsin: 15 additional data centers approved at former Foxconn site. Each one needs full networking buildout.
• 1.6T switch cycle launching late 2026 drives AEC replacement cycle — existing installations upgrade to higher-speed AECs.
• Azure demand exceeds supply through at least June 2026. $625B contracted cloud backlog.
• Three Mile Island nuclear restart (835MW) and $10B Brookfield renewable deal for long-term power.

CRDO revenue implications:

• Steady $65-75M/quarter base, growing toward $80-100M by FY27. Microsoft's power constraints actually benefit CRDO by smoothing out demand (no huge spikes, no huge dips).
• 1.6T upgrade cycle is a second demand driver on top of new builds. Existing Microsoft AEC installations from 2021-2024 will need higher-speed replacements.
• Maia 200 doesn't change AEC demand. Custom chips still need the same rack-to-switch connectivity. AEC demand is chip-agnostic.

→ Microsoft demand runway: Steady growth, power-paced. Expect $80-100M/quarter by FY27, $100-120M by FY28. The most predictable of the four customers.

Meta — The Sleeping Giant

Key dynamics:

• Prometheus is the catalyst. 500K GPUs, 1GW, three linked sites in Ohio. Coming online 2026. This is a Colossus-scale cluster — and it needs full networking buildout including AECs.
• NVIDIA multiyear deal (Feb 2026): "Millions" of Blackwell and Rubin GPUs, plus Spectrum-X networking. Spectrum-X = AEC opportunity. This is explicitly NVIDIA's Ethernet-based AI networking platform that uses AECs.
• Hyperion (Louisiana) is 5GW at full build. That's 2.5x xAI's entire Memphis complex. Multi-year buildout through ~2030.
• Meta purple cables confirmed. OCP Summit rack display. Already buying from CRDO at ~$33M/quarter.

CRDO revenue implications:

• **Meta at 33M/quartertodayiswherexAIwas2 − 3quartersago * *(20M in Q4 FY25). If Prometheus drives a similar surge, Meta could reach $80-150M/quarter by FY27-FY28.
• The NVIDIA Spectrum-X inclusion in Meta's deal is a direct AEC demand signal. NVIDIA's networking platform for AI uses AECs for short-reach. Every GB200 NVL72 rack Meta deploys needs them.
• Hyperion buildout (2027-2030) creates a 4+ year demand runway beyond Prometheus.
• Meta is the customer most likely to "xAI-ify" — going from modest to massive in 2-3 quarters. The Prometheus online date in 2026 is the trigger.

→ Meta demand runway: Accelerating. $33M today → $80-150M/quarter by mid-FY27 as Prometheus comes online. Hyperion extends the runway through FY30+. This is the next big ramp.

Google (Speculative 5th Customer)

No confirmed AEC relationship. Currently ~$8M/quarter. Google's TPU architecture uses custom interconnects (OCS + optical), which may limit AEC adoption vs NVIDIA-based hyperscalers. However, Google does deploy NVIDIA GPU clusters for some workloads, and those would use standard Ethernet networking with AECs.

→ Small but could grow if Google expands NVIDIA GPU deployments. Not a material driver near-term.

FY27-FY28 Revenue Build by Customer (Updated with Buildout Context)

Fleming guided FY27 to "nearly $2B" (>50% YoY from ~$1.33B). My per-customer build suggests $1.84-2.80B is the range, with $2.0-2.2B as the central case. The upside comes from Meta ramping faster than expected and ZF Optics exceeding "material component" guidance.

Key Takeaways for the CRDO Thesis

1. The infrastructure buildout is not peaking — it's accelerating.

• Combined CY2026 capex across just Amazon (200B), Microsoft(120-145B), Meta (115 − 135B), andxAI(30B+) = **465 − 510B + * * .Upfrom 345B in CY2025. That's 35-48% growth in infrastructure spend.
• These aren't plans. They're guided numbers from earnings calls.

2. Each customer is at a different stage of their buildout cycle.

• xAI: Mid-cycle (Colossus 2 installing, Building 3 converting)
• Amazon: Early-cycle on the CY2026 mega-ramp ($200B, doubling capacity)
• Microsoft: Steady expansion, power-gated
• Meta: Pre-ramp on Prometheus (the biggest catalyst ahead)
• This staggering means CRDO doesn't face a synchronized cliff. When xAI's Colossus 2 wave ends, Meta's Prometheus wave begins.

3. Custom silicon doesn't reduce AEC demand.

• AWS Trainium, Microsoft Maia, Meta MTIA — all custom chips. None of them change the rack-to-switch connectivity layer. AECs are chip-agnostic. Whether it's an H100, GB200, Trainium3, or Maia 200 inside the server, the cable connecting it to the network switch is the same AEC.
• This is perhaps the most underappreciated point. The "NVIDIA dependency" narrative misunderstands where AECs sit in the stack.

4. The demand staggering makes Fleming's "mid-single-digit sequential quarterly growth" conservative.

• If Meta ramps even moderately while Amazon stays steady and xAI maintains, the QoQ growth could exceed mid-single-digits for several quarters.
• Fleming's conservatism has been well-documented: guided $800M for FY26, delivering $1.33B.

Methodology Notes

• Disclosed concentration percentages from 10-Q filings and earnings call transcripts
• Cross-quarter linkages from CFO commentary (Q2 and Q3 calls)
• Customer identities matched using SEC filings, CEO conference statements, analyst notes
• Dollar amounts = percentage × total revenue (estimates where percentages are approximate)
• Q2 FY26 percentages use rolling tracker estimates (~30/25/20/12) which differ slightly from initial analysis table (42/24/16); the rolling tracker figures produce more consistent cross-quarter dollar progressions
• Earlier quarters (Q3 FY25 and prior) have less granular disclosure

-wsm (Long CRDO 11.0% equity + 2.8% LEAPS Jan'28 $115C)