AI infrastructure field guide · 19 July 2026

Inside NVIDIA
Vera Rubin.

The complete hardware supply chain and investment map behind NVIDIA’s next AI platform—covering HBM4, wafers, packaging, optics, power, racks, CPUs and networks.

Full-stackRubin GPU through Spectrum-6
HBM4the decisive memory layer
100+ kWrack-scale power challenge
28 namessuppliers + comparables screened
01 / Executive verdict

One platform.
Three kinds of exposure.

The cleanest investment mistake is to label every AI-infrastructure company a “Rubin supplier.” Some companies sell a part inside NVIDIA’s reference system. Others enable the surrounding datacenter. A third group merely competes for the same capital budget.

SK hynix should be in the primary list. It has a direct HBM4 relationship with NVIDIA, and its new Nasdaq ADR—SKHY—began trading on 10 July 2026. Its absence from an exchange-listed screen after that date is an omission, not a taxonomy choice.

Intel belongs, but only with a configuration label. DGX Rubin NVL8 uses two Intel Xeon 6776P host processors. The rack-scale Vera Rubin NVL72 instead centers NVIDIA’s own Vera CPU. Arm supplies architectural IP economics; NVIDIA designs Vera; TSMC manufactures the silicon.

Arista, Cisco and Marvell belong in the comparison, not automatically in the base bill of materials. Arista and Cisco address scale-out Ethernet and cluster operations. Marvell is strategically connected through custom silicon, optics and NVLink Fusion. NVIDIA’s reference rack still relies heavily on NVIDIA-designed NVLink, ConnectX, BlueField and Spectrum networking.

Bucket A · highest evidence

Confirmed content

A named component, manufacturing step or reference-system design win: NVIDIA, SK hynix, Micron, TSMC, Intel in NVL8, and the confirmed rack/power/optical manufacturers.

Bucket B · real but indirect

Ecosystem exposure

Products deployed around the rack or enabled through partnerships: Arista, Cisco, Marvell, Arm, server OEMs and optical vendors whose exact SKU content can vary.

Bucket C · budget comparison

Demand & alternatives

Hyperscalers and neoclouds buying accelerated infrastructure, plus AMD as a CPU/GPU comparison—not a confirmed Vera CPU supplier.

“Will beat” is not the same as “is forecast to grow.” The financial tables below use consensus estimates. A company beats only when reported results exceed the consensus that existed immediately before the release. Likewise, most issuers have not provided formal full-year 2027 guidance; “2027” is principally analyst consensus, not management guidance.
02 / The machine before the stocks

Rubin is not a GPU.
It is a system.

NVIDIA describes Rubin as a full-stack platform spanning compute, memory, scale-up fabric, scale-out networking, DPUs and rack-scale infrastructure. The economic value moves through multiple layers before the first token is generated.

The Rubin GPU performs accelerated compute and is paired with HBM4. The Vera CPU feeds and coordinates the system. NVLink 6 binds accelerators into a coherent scale-up domain. ConnectX-9 and BlueField-4 handle high-speed connectivity and infrastructure processing. Spectrum-6/Spectrum-XGS extends the fabric across racks. Power conversion, liquid cooling, copper, optics, PCBs, advanced packaging and final system integration make the architecture physically deployable.

This is why revenue attribution must be conservative. NVIDIA may name an ecosystem partner without disclosing its content per rack; an OEM may ship Rubin systems while capturing only a thin integration margin; an optical vendor may benefit from cluster scale without having a transceiver inside every reference rack.

ComputeRubin GPU · Vera CPU · optional x86 hostsNVDA designs; TSM manufactures; INTC wins DGX NVL8 host socket
MemoryHBM4 stacks, controllers and advanced packagingSKHY · MU memory; TSM packaging/foundry
Scale-upNVLink 6 switch fabric and copper connectivityPrimarily NVDA-controlled reference architecture
Scale-outConnectX-9, BlueField-4, Spectrum-6; Ethernet alternativesNVDA core; ANET, CSCO, MRVL as ecosystem/comparison
OpticalLasers, transceivers, fiber, connectors and manufacturingCOHR · LITE · GLW · FN
Physical plantPower shelves, UPS, cooling, racks and integrationVRT · DELL · HPE · SMCI · IBM
03 / Public-company component map

Who builds what—
and how certain is it?

U.S.-listed securities and Rubin relevance
Company / tickerHardware or roleClassificationEvidence boundary
NVIDIA
NVDA
Rubin GPU, Vera CPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-X/Spectrum-6Platform ownerDirect, disclosed NVIDIA products
SK hynix
SKHY
HBM4/HBM4E for Vera Rubin-class systemsDirectMulti-year NVIDIA memory partnership; newly Nasdaq-listed ADR
Micron
MU
HBM4 and advanced memoryDirect / qualifiedHBM supplier exposure; allocation can change by generation and customer
TSMC
TSM
Leading-edge wafer fabrication and advanced packagingDirectFoundry and packaging backbone
Vertiv
VRT
High-density power and liquid-cooling infrastructureDirect ecosystemReference-architecture collaboration; site content varies
Coherent
COHR
Lasers, optical components and transceiversOptical ecosystemCluster-level optical demand, not fixed content per rack
Corning
GLW
Optical fiber, cabling and connectivityDatacenter fabricBuildout beneficiary; deployment-dependent
Lumentum
LITE
Optical engines, lasers and datacom componentsOptical ecosystemAI optics exposure; exact Rubin content not universal
Fabrinet
FN
Precision manufacturing for optical communicationsManufacturingIndirect content through optical customers
Dell
DELL
Rubin-based systems, racks, storage and integrationOEMConfirmed NVIDIA infrastructure ecosystem
HPE
HPE
AI servers, liquid-cooled systems and servicesOEMConfirmed ecosystem; configuration varies
Supermicro
SMCI
Rack-scale servers and liquid-cooled integrationOEMSystem supplier, lower-margin integration economics
IBM
IBM
Enterprise AI infrastructure, software and integrationEcosystemNot a core Rubin semiconductor supplier
Intel
INTC
2 × Xeon 6776P host CPUs in DGX Rubin NVL8Direct in NVL8Do not generalize the socket win to Vera Rubin NVL72
Arm
ARM
CPU instruction-set/IP economics behind Arm-compatible Vera coresIP exposureNVIDIA designs Vera; Arm does not manufacture the CPU
Arista
ANET
Scale-out AI Ethernet switching and operationsAdjacentNot a confirmed component inside the reference rack
Cisco
CSCO
Enterprise AI networking; Cisco OS with NVIDIA Spectrum siliconPartnerReal NVIDIA partnership, principally cluster/enterprise layer
Marvell
MRVL
Custom XPUs, DSP/optical silicon and NVLink Fusion connectivityStrategic adjacentImportant future ecosystem role; not base NVL72 silicon
AMD
AMD
Alternative server CPUs and AI acceleratorsComparableNo confirmed CPU SKU in NVIDIA’s Rubin reference systems
04 / The CPU and network corrections

The labels matter.
The sockets are different.

Vera Rubin NVL72

NVIDIA Vera CPU

NVIDIA designs the Arm-compatible Vera CPU and couples it to Rubin at rack scale. TSMC supplies manufacturing. Arm receives IP-related economics, but does not build the physical processor.

DGX Rubin NVL8

Intel Xeon 6

NVIDIA’s published NVL8 configuration names two Intel Xeon 6776P host processors. This is a direct, configuration-specific design win and the strongest reason to include INTC.

Comparison only

AMD EPYC

AMD is relevant to AI-server capital allocation and x86 alternatives. Without a disclosed Rubin reference-system CPU SKU, it belongs in the comparison set rather than confirmed content.

Inside / close to rack

NVIDIA

NVLink, ConnectX, BlueField and Spectrum form the default networking stack.

Scale-out

Arista

High-performance Ethernet fabrics connecting AI clusters, with strong hyperscaler exposure.

Enterprise

Cisco

Networking, security and operations, including systems combining Cisco software with NVIDIA silicon.

Custom silicon

Marvell

DSPs, optical connectivity, custom accelerators and an expanding NVLink Fusion relationship.

05 / Consensus growth screen

Double-digit growth—
forecast, not promised.

The most useful screen is the next fiscal year over the current fiscal-year estimate. Fiscal calendars differ, and the figures are non-GAAP where the data provider reports adjusted EPS. These numbers answer “who is expected to grow?”—not “who will beat?”

14supplier / adjacent names with ≥10% revenue and EPS growth
4 of 5hyperscalers with double-digit revenue growth
3neoclouds with >75% expected revenue growth
1mature supplier below 10% on both: IBM
Supplier and adjacent-company consensus — current FY estimate to next FY estimate
TickerRevenue: current → nextRev growthEPS: current → nextEPS growthBoth ≥10%?
NVDA$393.48B → $559.35B+42.2%$8.99 → $12.83+42.7%YES
MU$129.66B → $234.11B+80.6%$73.39 → $150.77+105.4%YES
TSMNT$5.21T → NT$6.69T+28.4%NT$100.40 → NT$128.80+28.3%YES
SKHY₩358.97T → ₩535.21T+49.1%₩319.44K → ₩454.01K+42.1%YES
VRT$13.89B → $17.85B+28.5%$6.50 → $8.87+36.5%YES
COHR$7.06B → $9.57B+35.6%$5.45 → $8.24+51.2%YES
GLW$18.94B → $22.59B+19.3%$3.20 → $4.29+34.1%YES
LITE$2.99B → $5.60B+87.3%$8.23 → $18.30+122.4%YES
FN$4.60B → $5.66B+23.0%$13.81 → $17.26+25.0%YES
DELL$171.26B → $190.78B+11.4%$18.40 → $21.53+17.0%YES
HPE$44.94B → $49.90B+11.0%$3.41 → $4.00+17.3%YES
SMCI$39.68B → $52.44B+32.2%$2.59 → $3.12+20.5%YES
IBM$71.14B → $74.13B+4.2%$12.19 → $13.31+9.2%NO
INTC$58.86B → $66.02B+12.2%$1.10 → $1.58+43.6%YES
ARM$5.97B → $8.09B+35.5%$2.17 → $3.08+41.9%YES
ANET$11.62B → $14.38B+23.8%$3.64 → $4.47+22.8%YES
CSCO$62.91B → $68.79B+9.3%$4.28 → $4.79+11.9%EPS only
MRVL$11.53B → $16.72B+45.0%$4.05 → $6.20+53.1%YES
AMD$49.64B → $78.07B+57.3%$7.43 → $13.46+81.2%YES

Fastest paired growth

Among the supplier/adjacent set, the most aggressive consensus combinations are:

LITEMUAMDMRVLSKHYCOHR

Growth with direct content

The cleaner direct-exposure shortlist combines disclosed content and double-digit consensus growth:

NVDASKHYMUTSMVRTINTC*

*Intel content is specific to DGX Rubin NVL8.

Where the screen fails

IBM misses both thresholds; Cisco misses revenue; loss-making neocloud EPS growth is not economically comparable to profitable-company EPS growth.

06 / Price and forward multiples

Growth has a price.
Sometimes a very high one.

Prices are 17 July 2026 closes. Forward P/E and forward P/S are point-in-time consensus ratios checked 19 July. A low multiple can represent cyclicality or risk; a high multiple can represent quality, scarcity—or expectations with no room for error.

Rubin suppliers and adjacent comparables — market snapshot
TickerPriceForward P/EForward P/SInterpretive note
NVDA$202.8120.34×11.27×Platform economics plus exceptional expected growth
MU$848.955.92×4.14×Very low earnings multiple reflects memory-cycle skepticism
TSM$398.3718.49×N/M*Raw P/S vendor field is not ADR/currency normalized
SKHY$154.03≈5.8×*≈3.7×*Fresh ADR premium; normalized estimates remain approximate
VRT$289.5642.38×7.58×Premium valuation for power/cooling scarcity
COHR$277.6037.13×6.14×Optical-cycle acceleration priced in
GLW$154.6145.85×6.74×Fiber/AI rerating versus historical profile
LITE$732.8246.00×11.36×Fastest screen growth paired with one of the richest multiples
FN$478.3929.19×3.17×Manufacturing economics, customer concentration risk
DELL$396.3421.31×1.49×Low sales multiple reflects OEM margins
HPE$45.8211.90×1.25×Cheaper integrator exposure
SMCI$24.188.28×0.32×Low multiple carries execution/governance and margin risk
IBM$212.6717.49×2.79×Mature enterprise exposure; slower growth
INTC$95.0487.72×7.90×Recovery-year EPS makes P/E unusually sensitive
ARM$267.19123.24×47.68×Scarce IP asset with extreme expectations
ANET$168.6144.53×17.43×High-quality networking growth at premium price
CSCO$111.9423.86×6.52×Mature networking with less top-line torque
MRVL$188.6841.55×13.01×Custom-silicon/NVLink optionality priced aggressively
AMD$495.7656.08×14.43×Alternative accelerator growth, not Rubin content
ADR normalization: raw forward P/S fields for TSM and SKHY mix U.S. security prices with local-currency financial forecasts. SKHY is especially unstable because the ADR was only days old and traded at a premium to Seoul. The approximate SKHY ratios normalize the 1 ADR = 0.1 Korean common share conversion and current consensus; they should not be treated as precision values.
07 / Hyperscaler and neocloud comparison

The buyers—and
the leveraged buyers.

Suppliers monetize the build. Hyperscalers monetize workloads running on the build. Neoclouds often monetize scarce capacity but finance a much larger proportion of the asset base. Their revenue growth can be spectacular while EPS remains negative.

Demand-side comparison — price, valuation and next-year consensus
CompanyPriceFwd P/EFwd P/SNext rev growthNext EPS growth
Alphabet GOOG$346.1227.12×8.26×+19.8%+2.8%
Meta META$646.0119.82×6.21×+19.5%+13.7%
Amazon AMZN$247.2329.55×3.14×+13.0%+13.8%
Microsoft MSFT$393.8221.26×7.94×+16.7%+15.4%
Oracle ORCL$126.4115.71×4.06×+45.9%+35.7%
CoreWeave CRWV$73.21N/M2.55×+98.2%Loss narrows
Nebius NBIS$177.71N/M8.54×+236.6%Turns positive*
IREN IREN$33.62N/M6.05×+301.9%Loss narrows
Applied Digital APLD$25.79N/M12.93×+75.3%Loss widens

*Consensus EPS transition from a loss to a small profit; percentage growth is not meaningful across zero.

Best paired hyperscaler screen: Oracle has the fastest expected revenue and EPS growth in this comparison, followed by Microsoft, Meta and Amazon on the “both above 10%” test. Alphabet clears the revenue threshold but not the next-year EPS threshold. That is a growth screen—not an investment ranking.
08 / How to read the opportunity

Content, growth,
price, risk.

A useful Rubin thesis needs four independent answers. If one is missing, the conclusion is probably a narrative rather than analysis.

Content: Is the company’s product physically present, optionally present, or merely adjacent? Ask for the named SKU, reference design or partnership scope.

Growth: Is AI materially changing company-wide revenue and EPS, or only one segment? Small suppliers can have more percentage torque but much higher customer concentration.

Price: Does the forward multiple already assume several years of flawless execution? Compare P/E with margin structure and P/S with the gross margin required to justify it.

Risk: Memory is cyclical; optics can face inventory corrections; OEM revenue can be huge with thin margins; neocloud growth can require continuous debt and equity capital.

Allocation risk

A qualification does not guarantee volume. NVIDIA and customers can split HBM, optics and system orders among suppliers or redesign content over time.

Capex digestion

Hyperscaler spending can pause even when long-run AI demand remains intact. Suppliers experience the pause before software demand disappears.

Power constraint

Grid interconnects, substations and cooling can delay deployments after chips are ordered, shifting revenue between quarters.

Margin mix

Server OEMs can report explosive AI revenue without semiconductor-like gross margins. Revenue growth alone can overstate value capture.

Multiple compression

At 40–120× forward earnings, a company can hit estimates and still fall if the market lowers the multiple assigned to those earnings.

Estimate quality

Fresh ADRs, rapid-cycle businesses and loss-making neoclouds produce noisy consensus ratios. Treat false precision as a warning.

09 / Bottom line

The corrected list
is broader—but stricter.

SK hynix is a primary Rubin supplier. Intel is a direct NVL8 CPU supplier. Arm is an IP beneficiary. Arista, Cisco and Marvell are necessary comparison names—but their exposure sits mainly at the network and custom-silicon ecosystem layer, not automatically inside the standard NVL72 rack.

The strongest “direct content plus double-digit consensus growth” cluster is NVIDIA, SK hynix, Micron, TSMC and selected power/optical suppliers. The highest projected growth often comes with cyclicality, customer concentration or rich valuation. The cheapest multiples often carry recovery risk, governance concerns or structurally lower margins.

Among networking names, Arista offers the cleanest paired revenue/EPS growth profile, Marvell offers the strongest custom-silicon and NVLink optionality, and Cisco offers broader enterprise distribution with slower top-line growth. They are not interchangeable.

Among buyers, Oracle screens with the highest paired next-year growth, while the neoclouds deliver far faster revenue growth at the cost of negative earnings and financing intensity. The winner is not simply the row with the largest percentage.

10 / Sources and methodology

Evidence first.
Estimates dated.

Hardware classifications rely primarily on NVIDIA and supplier disclosures. Prices are 17 July 2026 closes. Consensus statistics were checked 19 July 2026 through StockAnalysis/S&P Global Market Intelligence. Fiscal periods follow each issuer; currencies remain local for TSM and SK hynix forecasts. Calculated growth equals next fiscal-year consensus divided by current fiscal-year consensus minus one.

  1. 01NVIDIA Rubin platform announcement
  2. 02NVIDIA DGX Rubin NVL8 specification
  3. 03NVIDIA HGX Rubin platform options
  4. 04SK hynix–NVIDIA memory partnership
  5. 05SK hynix Nasdaq listing
  6. 06SK hynix ADR registration
  7. 07Intel Xeon 6 in DGX Rubin NVL8
  8. 08Cisco–NVIDIA enterprise AI partnership
  9. 09NVIDIA Taiwan manufacturing ecosystem
  10. 10Market and valuation data methodology
  11. 11SK hynix consensus forecast data
  12. 12Example consensus forecast detail
Important: this is an analytical research article, not investment advice. Market prices and consensus estimates change continuously. “Direct” describes evidence of product/platform participation, not a forecast of revenue share. “Both ≥10%” is a mechanical screen, not a prediction that results will beat Wall Street estimates.