Picture a company with a working optical engine, a CXL memory controller, or a new cold plate. An engineer at a hyperscaler liked the demo. The board wants to know when revenue arrives. Everyone has read that AI data center spending is measured in trillions, so the question in the room is why the purchase order has not come.
The spending is real. Dell'Oro Group expects worldwide data center capital spending to pass $1 trillion in 2026 [1] and $3 trillion by 2030 [2]. But nobody sells to that number. A supplier sells to one platform program, at one account, through one system maker, inside one buying window. This guide covers how those decisions are made, what each buyer needs to see, and how to plan for a cycle that can take two years.
The Headline Number Is Not Your Market
The first mistake is treating total data center spending as the addressable market. It measures what operators spend on servers, networks, buildings, and equipment. The market for a specific product is a different figure, measured differently, and the two overlap. A switch sold to a hyperscaler shows up in that hyperscaler's capital spending and again in the switch market.
Keep three kinds of figure apart:
| Figure | What it measures | Example | Use it for |
|---|---|---|---|
| Operator spending | What hyperscalers, AI clouds, and operators invest in total | More than $1 trillion of data center capex expected in 2026 [1] | How much buying power an account has |
| Supplier revenue: memory | What vendors receive for high-bandwidth memory (HBM) | $54.6 billion estimated for 2026, up 58% [3] | Sizing a memory or packaging product |
| Supplier revenue: optics | What vendors receive for datacom optical components | More than $19 billion in 2025 [4] | Sizing an optical module or component |
| Supplier revenue: switching | Sales of AI back-end switches, scale-up and scale-out | About $1 trillion from 2026 to 2030 combined [5] | Sizing switch silicon, systems, and interconnect |
| Supplier revenue: physical layer | Power, cooling, and rack equipment from manufacturers | $120 billion forecast for 2030 [6] | Sizing power and cooling equipment |
Never add these together. They answer different questions, and summing them counts the same dollar several times.
Your real market is narrower and more useful: the accounts whose next platform generation needs what your product does, multiplied by the share of that platform your product can occupy, inside the window when the architecture is still open. That is a number you can defend to a board, and it tells you which five accounts matter.
How Hyperscalers Actually Choose Components
Enterprise selling starts with a request for proposal. Hyperscaler selling usually does not. McKinsey's 2026 analysis of data center suppliers found that hyperscalers choose suppliers based on whether they can show a roadmap aligned with the future architecture, and that requirements are set by chip makers' reference architectures before any formal request goes out [7]. By the time a procurement event is visible, the architecture is often already decided.
Three more facts shape the sale:
- The hyperscaler chooses and the manufacturer orders. Astera Labs, which sells connectivity chips into AI servers, reports that hyperscalers and system makers often dictate sourcing decisions, while contract manufacturers and design partners build the systems and place the orders [8]. You need the engineering decision and the manufacturer relationship.
- The cycle is long. The same filing describes up to two years from a design win to volume shipments, after a technical evaluation and a qualification that can take several months or more [8].
- Open specifications decide who gets a seat. AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI formed the OCI MSA to write an open specification for optical scale-up links, with a multi-vendor supply chain as the stated goal [9]. Buyers want parts they can source from more than one supplier, and the specification work is where that is decided.
For physical-layer suppliers, delivery is now part of the product. McKinsey puts North American lead times at about 80 weeks for medium-voltage switchgear and about 50 weeks for transformers [7]. A supplier that can commit to a delivery date has an argument a cheaper competitor may not.
- $1T+ data center capex expected in 2026
- Operator spending, not any one product's market: use it to rank accounts, not to size a product
- [1]
- Up to 2 years from design win to volume
- Qualification alone can take several months or more, so plan the runway around the full cycle
- [8]
- About 80 weeks for medium-voltage switchgear
- North American lead times make delivery certainty a buying criterion for power equipment
- [7]
- $19B+ datacom optical components in 2025
- A measurable optics market today, with 1.6T modules ramping
- [4]
- $54.6B estimated HBM market in 2026
- Memory is a large supplier market of its own, not only a GPU part
- [3]
The Five Gates to a Design Win
Every hyperscaler design win passes the same five gates. Each is a question the buyer is really asking, and a product that cannot answer one stops there however good it is.
| Gate | What the buyer is really asking | Evidence that gets you through |
|---|---|---|
| Architecture fit | Does this belong in our next platform generation? | A roadmap mapped to the accelerator, network, and memory plans, and a clear view of the competing options |
| Technical proof | Is the advantage real at the system and workload level? | A benchmark on the buyer's workload, a cost of ownership model, and watts per bit or per terabyte, latency, and reliability data |
| Ecosystem fit | Will our chip, system, software, and supply partners support it? | Support in a reference architecture, a standards plan, and the system makers who will build it |
| Supply qualification | Can this company make, deliver, and support it at our scale? | A capacity plan, a qualification roadmap, and a second source |
| Commercial adoption | Is the value large enough to justify switching? | An account-specific return on investment and a pilot-to-production plan |
Most early-stage suppliers over-invest in the second gate and under-invest in the third and fourth. A strong benchmark does not survive a sourcing review that finds a single fab, no second source, and no system maker ready to integrate the part.
A Hyperscaler Account Is Many Teams
A single enthusiastic engineer is not an account. The decision runs through technical sponsors who own a problem today, architecture owners planning the next generation, platform teams who will integrate and operate the part, sourcing and supply chain teams who care about price, capacity, and a second source, partner organizations who want support from the chip and system makers they already use, and executives who weigh cost and risk against capacity plans.
Each team has to agree with the same quantified case. When they hear different stories, the program stalls in the gap between them.
Measure Progress by Evidence, Not Meetings
Pipeline reviews that count meetings reward activity. Track each account through stages that each require proof:
- 1Market identified. You have a product-level market size and know the architecture window.
- 2Account qualified. Spending, roadmap, and timing signals point to a real program.
- 3Technical sponsor. A named engineering team is evaluating the problem and your product.
- 4Ecosystem aligned. The chip, system maker, software, or standards relationships you need are in place.
- 5Benchmark or pilot. A proof on the customer's workload or system is under way.
- 6Qualification. Reliability, interoperability, production, and supply are being validated.
- 7Design win. Your product is specified in a platform or deployment.
- 8Expansion. More sites, generations, regions, or ecosystem customers are added.
An account moves forward only when the evidence for the next stage exists. That one rule makes forecasts honest and shows the board exactly where each account is stuck.
Your First Customer Does Not Have to Be a Hyperscaler
The four largest clouds are not the only buyers. Dell'Oro reported that neocloud and AI model builder capital spending grew fastest of any customer segment in the second quarter of 2026 [10], and CoreWeave reported a contracted backlog of about $104 billion in August 2026 [11]. AI clouds, developers, and colocation providers often move faster than the largest clouds, and a production deployment with one of them gives you the reference that makes the hyperscaler conversation easier.
What to Prove, by Kind of Supplier
The gates are the same for everyone, but the evidence differs:
- Photonics. Where the product lives (scale-up, scale-out, memory fabric, or between data centers), at what speed and reach, and what it improves in watts per terabit, cost per gigabit, density, or reliability. Our photonics go-to-market guide walks the optical supply chain from laser to operator.
- Memory and CXL. How much HBM, DRAM, or accelerator capacity the product saves or puts to better use, and which workloads accept the added latency. The AI memory wall works through the numbers for a 405 billion parameter model.
- Power and cooling. Rack density supported, heat removed, megawatts delivered, lead time, and how the equipment is serviced. The power and cooling guide covers who specifies it during facility design.
Questions Suppliers Ask
How long does it take to get designed in at a hyperscaler? Plan for up to two years from a design win to volume shipments, with qualification alone taking several months or more [8]. The architecture conversation that leads to the win usually starts earlier still.
Can we sell hardware through AWS Marketplace? No. AWS Marketplace lists software, data, machine learning models, and professional services, not physical components [12]. Components are chosen by the hyperscaler's own engineering and sourcing teams after a technical evaluation [8].
Do we need to join standards bodies? If your product category is being specified, yes. Buyers want multi-vendor supply, and specifications such as the OCI MSA are where the requirements are written [9]. Being in the room is how you learn what will be required before the request goes out.
How do we know which accounts to pursue first? Start from signals: which operators are raising capital spending, which are hiring the engineers who would evaluate your product, and which are active in the standards your product depends on. The Prospectory AI Infrastructure Leaderboard tracks those signals across hundreds of companies [13]. Treat a signal as a reason to look closer, not as proof of a buying program.
What to Do This Quarter
Pick three accounts, not thirty. For each one, write down the platform program and generation your product targets, the named people in each buying team, and the one benchmark that would move the technical sponsor. Then name the system maker or contract manufacturer who would build it in, and the specification your product has to meet.
If any of those blanks is empty, that is the work for the next ninety days. Tactical Edge helps infrastructure suppliers fill them: sizing the real market, ranking accounts, mapping the buying team, building the proof, and supporting the sale through to a design win. Our team has led AI infrastructure go-to-market inside AWS, including the Inferentia and Trainium accelerators, and worked on optical interconnect and CXL memory products. See how we work with suppliers, or start with the AI infrastructure market map to see where the spending goes.
References
[1]Dell'Oro Group, "AI Infrastructure Buildouts and Memory Cost Inflation Drove Data Center Capex Higher in 1Q 2026," June 2026. https://www.delloro.com/news/ai-infrastructure-buildouts-and-memory-cost-inflation-drove-data-center-capex-higher-in-1q-2026/
[2]Dell'Oro Group, "AI Buildout Maintains Momentum as Data Center Capex Surpasses $3 Trillion by 2030," August 2026. https://www.delloro.com/news/ai-buildout-maintains-momentum-as-data-center-capex-surpasses-3-trillion-by-2030/
[3]SK hynix Newsroom, "2026 Market Outlook: Focus on the HBM-Led Memory Supercycle," January 2026. https://news.skhynix.com/en/2026-market-outlook-focus-on-the-hbm-led-memory-supercycle/
[4]Cignal AI, "Datacom Optical Component Revenue Surpasses $19B in 2025," April 2026. https://cignal.ai/2026/04/datacom-optical-component-revenue-surpasses-19b-in-2025/
[5]Dell'Oro Group, "AI Back-end Switch Sales to Approach $1 Trillion over the Next Five Years," July 2026. https://www.delloro.com/news/ai-back-end-switch-sales-to-approach-1-trillion-over-the-next-five-years/
[6]Dell'Oro Group, "Data Center Physical Infrastructure Market Forecast to Reach $120 Billion by 2030," August 2026. https://www.delloro.com/news/data-center-physical-infrastructure-market-forecast-to-reach-120-billion-by-2030/
[7]McKinsey and Company, "The $7 Trillion Data Center Build-Out: How Industrials Can Capture Their Share," March 2026. https://karriere.mckinsey.de/industries/technology-media-and-telecommunications/our-insights/the-7-trillion-dollar-data-center-build-out-how-industrials-can-capture-their-share
[8]Astera Labs, "Annual Report on Form 10-K for 2025," SEC filing, February 2026. https://www.sec.gov/Archives/edgar/data/1736297/000173629726000010/alab-20251231.htm
[9]TelecomTV, "Optical Scale-up Consortium Established to Create an Open Specification for AI Infrastructure," March 2026. https://www.telecomtv.com/content/next-gen-telco-infra/optical-scale-up-consortium-established-to-create-an-open-specification-for-ai-infrastructure-55073/
[10]Dell'Oro Group, "Data Center Capex Grew 92 Percent in 2Q 2026, Driven by Surging AI Demand and Memory Costs," September 2026. https://www.delloro.com/news/data-center-capex-grew-92-percent-in-2q-2026-driven-by-surging-ai-demand-and-memory-costs/
[11]CoreWeave, "CoreWeave Reports Strong Second Quarter 2026 Results," August 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx
[12]Amazon Web Services, "AWS Marketplace Product Types," AWS documentation. https://docs.aws.amazon.com/marketplace/latest/buyerguide/buyer-product-types.html
[13]Prospectory, "AI Infrastructure Leaderboard," sources dated October 16, 2026. https://prospectory.ai/ai-infrastructure-leaderboard/