Everything About AI
NextFin News - Alphabet and Blackstone's $5 billion AI infrastructure joint venture, announced in May with a target of 500 megawatts of compute capacity online by 2027, has hit delays at key data-center sites intended to host Google's Tensor Processing Units. The setback arrives less than four months after a deal that was marketed as a template for how deep-pocketed capital could clear the path for AI expansion — and it lands at a moment when roughly a third of the global data-center capacity slated for 2026 is actually under construction. The constraint is no longer chips, and it is no longer money. It is the power grid.
The venture, known internally as Project Braid, was designed to rent Google's AI processors to customers starting in 2027, joining a swelling class of "neoclouds" that lease out chips and capacity rather than selling a full cloud stack. Blackstone committed $5 billion in equity capital from funds it manages, one of the largest single bets an alternative asset manager has placed on the AI buildout. Google supplies the hardware — its custom-built TPUs — along with software and services. The structure was meant to split the problem neatly: Blackstone brings the real estate, energy relationships, and balance-sheet patience; Google brings the silicon that AI labs are lining up to access.
Now the newest, best-capitalized entrant into AI infrastructure is running into the same wall as everyone else. The companies have not disclosed how long the slippage runs or which sites are affected. What the silence underscores is that even the best-funded players cannot buy their way past the grid.
The Constraint Moved From Chips to the Grid
For the first wave of the AI boom, the binding constraint was obvious: who could get GPUs. Nvidia's accelerators were allocated months in advance, and access to chips defined who could compete. That bottleneck has not disappeared, but a second one has moved in front of it. A chip that cannot be plugged in is inventory, not infrastructure.
The data are stark. Sightline Climate's 2026 Data Center Outlook tracks at least 16 gigawatts of capacity slated to come online globally this year across roughly 140 projects — nearly triple the level built the previous year. Only about 5 GW is under construction. The other 11 GW sits at "announced," with no visible site progress despite typical build timelines of 12 to 18 months. Sightline expects 30% to 50% of the 2026 pipeline to be delayed or canceled. The reasons cluster in one place: grid interconnection queues that stretch three to seven years in major markets, shortages of transformers and switchgear, permitting gridlock, and growing local opposition.
That is why the Google-Blackstone delay is not an isolated execution hiccup. It is the clearest signal yet of a sector-wide transmission mechanism: capital can buy land and order chips in weeks, but it cannot buy a grid connection on demand. Hanwha Data Centers reports that 84% of data-center decision-makers now rank power availability among their top three site-selection criteria — ahead of traditional factors like fiber access or proximity to users. The industry has effectively admitted the queue itself is the bottleneck: PJM and ERCOT, the two largest U.S. grid operators, are rolling out interconnection reforms in 2026 because the old process could not absorb the load.
"We see a generational opportunity to invest capital at scale building AI infrastructure," Jon Gray, Blackstone's president and chief operating officer, said when the venture was announced. "This new company has enormous potential as it helps to meet the unprecedented demand for compute."
Four months later, "unprecedented demand" has met a very physical constraint. The question for investors is whether this is a cyclical shortage that will mean-revert or a structural regime shift that will not.
Why This Is Structural, Not Cyclical
The evidence points to structural — at least for the investment horizon that matters to a venture underwriting 2027 revenue. A cyclical claim would need three things: a short-term driver, a demonstrated mean-reversion pattern, and historical cycles that resolved on their own. None of those hold cleanly here. Grid interconnection is not an inventory cycle; it is a permitting and physics problem. Transmission lines take a decade to site and build. Transformers are bespoke, high-voltage pieces of equipment with multi-year lead times and a concentrated global supplier base. When the grid operators themselves are rewriting their interconnection rules because the queue became the bottleneck, that is a system admitting it was not built for this load.
The adaptation argument is real and deserves weight. Developers are moving toward on-site generation — natural gas turbines, behind-the-meter solar paired with storage — that bypasses the queue entirely. Some operators have signed dedicated power deals. If enough capacity comes online off-grid, the constraint loosens. But that adaptation has its own limits. On-site generation works for sites with fuel access and emissions headroom; it does not scale to the tens of gigawatts the industry has announced. And it shifts the bottleneck rather than removing it: now the constraint is gas pipeline capacity, turbine supply, and local air permits. The form of the bottleneck changes; the fact of the bottleneck does not.
So the call is structural for the next several years. Capital-intensive players with long-duration funds and real-estate expertise — Blackstone chief among them — are better positioned than leveraged neoclouds to wait it out. But "better positioned" is not the same as "unaffected."
The Second-Order Squeeze on Neocloud Economics
Here is the implication the market is not fully pricing. A neocloud's business model is a race against depreciation. GPUs and TPUs lose economic value every month they sit idle, while the debt or equity that bought them accrues cost from day one. CoreWeave, for example, has reported roughly 3.1 gigawatts of contracted power against about 850 megawatts actually active — a gap that looks like optionality until the revenue clock starts ticking. Lambda, a private neocloud, operates 15 to 20 U.S. data centers with more than 320 MW committed and is targeting 3 GW by 2030. Nebius is aiming for 800 MW to 1 GW connected and 2.5 GW contracted by the end of 2026.
When deployment slips, three things happen at once: revenue is pushed back, the hardware depreciation clock keeps running, and the cost of capital rises because lenders start underwriting counterparty risk. Smaller neoclouds carry customer-concentration risk on top of that — a handful of anchor tenants, contracts that may run shorter than the financing on the GPUs. Global Data Center Hub frames the underwriting rule plainly: always separate contracted power from deployed GPUs, because the two diverge widely.
The Google-Blackstone venture was supposed to be insulated from exactly this. Blackstone's $5 billion is patient equity, not leveraged bridge finance, and Google controls the chip supply. Yet the delay shows that even that fortress balance sheet cannot accelerate a grid connection. The second-order effect is a re-ranking of the neocloud hierarchy: the winners will be whoever owns contracted, deliverable power — not whoever has the biggest chip allocation.
Google's Coopetition Bet Gets Tested
There is a second tension inside this deal that the delay exposes. Google is effectively funding a competitor to its own Google Cloud. The logic is a dual-brand play: Google Cloud keeps the enterprise customers who want the full integrated stack, while the joint venture captures the "raw compute" segment that fears vendor lock-in and might otherwise drift to CoreWeave or AWS. Thomas Kurian, Google Cloud's chief executive, framed it at the launch as giving customers "more options for organizations to access accelerated compute capability."
That strategy only works if the new venture can actually deliver capacity on schedule. A delayed neocloud does not just lose its own customers — it hands them back to Google Cloud or to rivals. For Alphabet, which reported second-quarter revenue of $119.8 billion, up 24% year over year, with Google Cloud revenue surging 82% to $24.8 billion, the joint venture was a way to expand TPU reach without loading more capital expenditure onto its own balance sheet. That matters because Alphabet's capital spending is already under investor scrutiny: second-quarter CapEx reached $44.9 billion, exceeding operating cash flow of $39.1 billion and producing negative free cash flow of about $5.9 billion. On the July earnings call, management raised full-year 2026 capital expenditure guidance to a range of $195 billion to $205 billion, up from the prior $180 billion to $190 billion.
So the venture serves two masters: it is both an offensive move to capture compute demand and a defensive move to keep CapEx off Alphabet's books. Delays weaken both. If the 500 MW target slips, Alphabet either absorbs the delay in its own AI rollout or finds that the off-balance-sheet path is slower than building itself.
Blackstone's Wager: Scale Against a Broken Queue
The delay also tests Blackstone's broader thesis that it can out-build the bottleneck. The firm is not new to data centers: it owns QTS, one of the largest U.S. hyperscale operators, and in 2026 agreed to acquire AirTrunk, the largest data-center platform in the Asia-Pacific region, in a deal valued at roughly A$24 billion. AirTrunk brings more than 800 MW of capacity committed to customers and developable land that can support over 1 GW of future growth across Australia, Japan, Malaysia, Hong Kong, and Singapore. Before that deal, Blackstone said its data-center portfolio — including facilities under construction — stood at about $70 billion, with more than $100 billion in prospective development pipeline.
That scale is the point. A $5 billion venture is large for a single deal but small relative to a portfolio that spans tens of billions. Blackstone's argument is that a diversified, deep-pocketed owner can absorb a delayed site better than a neocloud that has staked its survival on one campus and one customer. It can sequence projects, shift capital across jurisdictions, and lean on existing utility relationships. It can also walk away from bad sites without jeopardizing the franchise.
But scale cuts both ways. A $100 billion prospective pipeline is only valuable if power can be delivered to it. If interconnection queues remain multi-year, a large share of that pipeline is optionality, not backlog. The Google-Blackstone delay is a reminder that in this market, announced capacity and energized capacity are two different currencies. Investors in alternative asset managers are starting to price that difference: Blackstone's shares last closed at $136.15, down about 28% from their 52-week high of $190.09 and down roughly 5% over the prior five trading days.
The Counter-Thesis — and What Would Prove It Wrong
The strongest case against reading too much into this is that delays are front-loaded, and the industry is adapting faster than the grid. Interconnection reforms in PJM and ERCOT are live in 2026. On-site generation is moving from pilot to standard practice. Blackstone is the largest global provider of data centers and has the real-estate and energy muscle to navigate permitting better than a startup. Core Scientific's stumbles were one operator's problem, not proof of a sector-wide dead end. Under this view, the Google-Blackstone slippage is a first-project learning curve, and the 2027 target still holds.
That is a fair argument, and it is why this is not a call to write off the venture. But it rests on a specific, testable premise: that the first 500 MW comes online on schedule in 2027. The falsifying signal is straightforward. If the venture announces a material push in the 500 MW timeline — the first phase moving into 2028 — the structural-bottleneck thesis is confirmed and the neocloud re-rating accelerates. Conversely, if the venture hits 500 MW on time, the adaptation thesis wins and the delays were indeed front-loaded noise. Watch the next milestone announcement from the venture on energized capacity. That single data point separates a sector-wide constraint from a single-project stumble.
What to Watch: Base, Upside, and Downside
The near-term read is uncomfortable for anyone long the neocloud trade: delays mean revenue pushes, margin pressure, and a higher bar for financing. The medium-term read is more selective: capital-rich, power-secure players — Blackstone, CoreWeave, the hyperscalers — separate from leveraged, power-insecure ones. The long-term read is structural: AI infrastructure is now an energy-and-permitting business as much as a technology business, and valuation multiples that priced it as pure software-adjacent growth will need to absorb industrial-cycle risk.
Who benefits: alternative asset managers with long-duration capital and existing data-center platforms; grid-equipment suppliers; on-site power developers. Who is exposed: neoclouds with contracted-but-not-energized capacity and short-dated customer contracts; any hyperscaler whose AI roadmap assumes on-time third-party capacity.
Base case: the Google-Blackstone venture reaches a scaled-back first phase in 2027, with the full 500 MW slipping modestly. Upside case: interconnection reforms bite faster than expected and on-site generation closes the gap, bringing capacity online near target. Downside case: permitting and transformer lead times push the first phase into 2028, triggering a broader re-pricing of neocloud timelines.
On the tape, shares showed no confirmed material move tied specifically to the delay report. Alphabet last closed at $342.48, inside a 52-week range of $224.79 to $408.61. The muted reaction is itself a data point: the market has not yet decided whether this is a single-project stumble or the first visible crack in the AI infrastructure thesis.
The AI boom was supposed to be won by whoever had the most chips. It is increasingly clear it will be won by whoever has the most electrons.
Explore more exclusive insights at nextfin.ai.
