Lambda has signed 320 MW and is targeting 3 GW by 2030. CoreWeave operates 250,000+ GPUs across 32 data centers. The operators getting infrastructure allocations in 2026 committed 12-24 months ago. Here is how that reservation stack works and what it leaves for everyone else.
Aerial view of a large-scale data center campus under construction rows of cooling units and server halls from above
Syaala Research
Infrastructure Market Analysis
Share
"87% of capacity under construction is already pre-leased before the building is delivered. The reservation clock started before most operators began their search."
Cushman and Wakefield North America Data Center Report, 2026
The reservation stack
Cushman and Wakefield reported 87% of capacity under construction was pre-leased before delivery in 2026. JLL put primary North American market vacancy at 1% at year-end 2025. These two numbers describe the same structural condition from different angles: the supply side of the data center market doesn't operate on a first-come, first-served basis. It operates on a commitment hierarchy that was established months or years before a building is complete.
The hierarchy has a clear shape. Hyperscalers sit at the top Google, Microsoft, Meta, Amazon with 10-15 year lease commitments, investment-grade balance sheets, and the negotiating position that comes from being a developer's anchor credit. Below them sit the large neoclouds: Lambda, CoreWeave, Nebius, Nscale, FluidStack, Crusoe. These operators are committing to multi-year, multi-MW reservations against capacity that doesn't exist yet. Then come creditworthy mid-market operators who can demonstrate lease commitment depth above the 10 MW threshold. Below that, in the thinnest part of the stack, is everyone else.
The distinction between tiers isn't just about size. It's about the terms each tier can credibly commit to. Hyperscalers and large neoclouds are signing 10-15 year leases at volumes that justify a developer breaking ground before a single rack is operational. That commitment depth is the actual currency of allocation. Without it, operators are competing for what's left after the reservation stack settles.
How the allocation hierarchy works in 2026
1- Hyperscalers - anchor pre-commitment
Google, Microsoft, Meta, Amazon. 10-15 year leases at 50-200+ MW per campus. Investment-grade credit. These deals are what justifies a developer's $300-400M construction commitment. Supply is de facto reserved at this tier before a building is announced.
2-Large neoclouds - forward reservation
Lambda, CoreWeave, Nebius, Nscale, FluidStack, Crusoe. Multi-year commitments of 10-50+ MW per site, signed 12-24 months ahead of need. GPU reservation volume and long-term customer contracts provide the revenue visibility developers require. Not always investment-grade credit, but commitment depth compensates.
3 Creditworthy mid-market - threshold access
Enterprise tenants and operators above the 10 MW investment-grade threshold. Access depends on lease term flexibility and credit rating. Supply at this tier is available but limited - it's the residual after the first two tiers absorb new construction.
4 Everyone else - secondary market and alternatives
The 1-10 MW segment without investment-grade credit or long lease commitment. Primary market new supply is structurally unavailable. Realistic options are secondary market inventory, prior-cycle buildings, and purpose-built midmarket capacity with different financing structures.
Lambda's 320 MW and what it means structurally
Lambda Labs announced 320 MW of signed capacity commitments in 2025, with a stated target of 3 GW by 2030. That target, if achieved, would make Lambda one of the largest GPU cloud infrastructure operators globally. More interesting than the headline number is the structure underneath it.
Bloomberg reported that Microsoft has been renting GPU capacity from Lambda as part of Microsoft's broader $60B+ cloud infrastructure commitment. That detail describes something unusual: a hyperscaler with its own massive data center footprint sourcing GPU compute from a neocloud. It indicates that GPU cluster optimization, network topology for AI training, and the operational overhead of maintaining high-density GPU infrastructure have created enough specialization that even a hyperscaler finds it efficient to outsource specific workloads.
This separation of concerns matters for how the infrastructure reservation stack reads. Lambda isn't competing with Microsoft for data center capacity. Lambda is providing a layer of GPU infrastructure that Microsoft can consume without operating it. The infrastructure demand signal flows through Lambda into the data center market where Lambda shows up as a large creditworthy tenant, not a hyperscaler.
That's the mechanism by which neocloud reservation activity tightens supply for everyone below them. Lambda's 320 MW of signed commitments are converting into developer ground-breaks. Each of those campuses absorbs a block of power and space that is now unavailable to the next tier of operator. The reservation stack advances by capacity, not by time.
Operator Scale Indicator Reservation Approach Notable
Lambda 320 MW signed, 3 GW 2030 target Multi-year forward commitments Microsoft renting capacity (Bloomberg)
CoreWeave 250,000+ GPUs, 32 data centers Lease + own build mix Multi-billion dollar customer contracts
Nebius Multi-year GB200 NVL reservations Forward capacity blocks European and US expansion
Nscale Multi-year GPU reservations Forward capacity blocks UK and continental expansion
FluidStack Aggregated multi-datacenter GPU pool Distributed reservation model Cross-market aggregation
Crusoe Own-build and lease mix Stranded power + colocation Waste-gas and curtailment power focus
The credit and scale filter
The 10 MW investment-grade threshold isn't an arbitrary number. It emerges from the economics of data center financing. A utility-scale campus today carries $300-400M in construction cost. The debt portion of that capital stack, particularly the insurance and pension debt that now funds the majority of large data center construction, requires Moody's Long-Term Credit Tenancy compliance an investment-grade anchor tenant on a lease that extends past the loan maturity date.
Data Center Knowledge described investment-grade credit as "the price of admission" for hyperscale-adjacent colocation allocations in Q1 2026. That framing is accurate but understates the compounding effect. It's not just a credit test. It's a credit test combined with a size test combined with a term test. An operator needs to be large enough to move the needle on a developer's pre-leasing percentage, creditworthy enough to satisfy debt underwriting, and willing to commit to a lease term that matches the financing structure. These three conditions together filter out most of the market.
Neoclouds clear this filter not primarily through balance sheet credit ratings most don't carry rated investment-grade credit in the traditional sense but through the revenue visibility of their customer commitments. A GPU cloud operator with multi-year, multi-hundred-million-dollar contracts from enterprise AI customers can demonstrate rent service coverage that satisfies a developer's underwriting requirements even without an S&P rating. The filter is ultimately about predictable cash flow, and large neoclouds have built that case through customer contract depth.
The 10-15 year standard lease term compounds the access problem. Most enterprise technology companies don't make infrastructure commitments on that timeline. A 10-year colocation lease signed today locks in terms through 2036 beyond a typical technology refresh cycle, beyond a typical PE hold period, well beyond the planning horizon most IT organizations operate on. The neoclouds signing these leases are betting their business on sustained GPU demand for a decade. Most operators aren't in a position to make that bet.
320 MW
Lambda signed capacity (Lambda announcements)
81.5%
Share of new supply preleased before delivery (Cushman and Wakefield)
1%
Primary North American market vacancy (JLL year-end 2025)
$73B
New data center debt financing in 2025 (JLL)
What's left for the 1-10 MW operator
The realistic supply options for an operator below the reservation stack threshold fall into three categories. None of them are large. Understanding which one applies to a given situation is more productive than trying to access primary market supply that's structurally unavailable.
Secondary markets with older inventory
Data centers in secondary markets - Atlanta, Dallas, Phoenix, Chicago, Denver that didn't achieve full hyperscale pre-leasing carry more accessible vacancy than primary markets. Construction debt on older buildings may be retired, removing the LTCT compliance requirement and enabling shorter lease terms. The tradeoff is real: older power infrastructure, lower supportable rack density, and in some markets, longer fiber paths to major network exchanges. But the access terms are different. Operators with flexibility on vintage and location find this pool more navigable than primary market RFPs.
Demand aggregation and shared reservation
Some operators are exploring demand aggregation pooling committed capacity requirements across multiple tenants to cross the threshold a developer will recognize. A consortium of 1-3 MW operators with combined demand of 10+ MW can, in theory, present a developer with the pre-leasing percentage that justifies construction. The execution is harder than the concept: operators need compatible power density requirements, compatible timeline, and willingness to negotiate collective lease terms. It's not a common path, but it's one that some regional infrastructure developers are actively facilitating.
Modular and factory-built alternatives
Factory-built modular data centers have a different financing structure than site-built facilities. Because the unit cost is lower and the deployment timeline is compressed 3-4 months from order to operational versus 18-24 months for conventional construction the capital commitment required to serve a 1-5 MW operator is achievable without the insurance debt that mandates LTCT compliance. The tradeoff is physical: modular facilities have lower campus density and in some configurations, higher cooling PUE than purpose-built hyperscale facilities. For AI workloads requiring high rack density, the latest generation of liquid-cooled modular infrastructure closes much of that gap.
The GPU cloud pricing data frames the economic stakes clearly. Lambda's B200 on-demand is at $3.49/hr. H100 cloud pricing ranged from $1.49 to $6.98/hr across providers in Q1 2026 per IntuitionLabs. That revenue per GPU-hour means a 1,000-GPU deployment generates between $1.5M and $7M in annual revenue at current market rates before utilization discounts and long-term contract pricing. An operator capturing even mid-market utilization on a 1,000-GPU cluster has the cash flow to service a meaningful infrastructure commitment. The barrier isn't revenue. It's access to the right supply tier, on terms that match the operator's lease commitment capacity.
The operators clearing the reservation stack in 2026-27 aren't necessarily larger or more creditworthy than those who aren't. They committed earlier. Lambda's 320 MW didn't get signed in 2026 those commitments started building in 2024. CoreWeave's 32 data centers represent years of site selection, lease execution, and infrastructure buildout that started well before the current AI demand wave hit mainstream visibility. The reservation clock runs ahead of the deployment clock by 12-24 months. That gap is structural, not temporary.
Sources
Lambda Labs capacity announcements - 320 MW signed capacity, 3 GW 2030 target
Bloomberg - Microsoft renting GPU capacity from Lambda, part of $60B+ cloud commitment
Cushman and Wakefield North America Data Center Report 2026 - 81.5% pre-leased before delivery, 87% preleased before delivery figures
JLL Year-End 2025 Data Center Report - 1% primary North American market vacancy
JLL Global Data Center Investment 2025 - $73B debt financing figure
Data Center Knowledge Q1 2026 - investment-grade credit as price of admission framing
IntuitionLabs Q1 2026 GPU Cloud Pricing Survey - H100 pricing range $1.49-$6.98/hr
CoreWeave public disclosures - 250,000+ GPU count, 32 data center footprint
Top comments (0)