Nvidia holds $279 billion in outstanding supply obligations for AI hardware that depends on data centers being physically built and permitted. David Sacks, who served as the first White House AI and Crypto Czar before returning to private investment, told listeners of the All-In Podcast released November 7 that by 2028, "the pipeline will be closed or significantly reduced." He isn't talking about demand weakening. He's talking about permits.
The warning landed as chip stocks posted their worst month in years: Nvidia fell 21.4%, Broadcom dropped 22%, Arista Networks lost 20.86%, and Vertiv shed 25.81% in the four weeks following the midterm elections. Infrastructure landlords held steadier: Equinix fell 1.85%, Digital Realty slipped 3.47%. That split tells you which end of the data center supply chain investors now believe is most exposed to what Sacks is describing.
What $279 Billion in Supply Obligations Assumes About Permits
Nvidia's most recent earnings period put data center revenue at $89 billion, up 117% year over year, with the next quarter guided to $108 billion, plus or minus 2%, according to Nvidia's official results. Those numbers sit on a foundation of $279 billion in outstanding supply obligations recorded on the company's balance sheet, representing orders customers have placed but not yet received. Every GPU in that backlog requires a powered, cooled, and physically permitted data center shell to plug into. The GPU itself is useless until the facility is ready.
Broadcom's CEO made the same dependency explicit on a recent earnings call. He stated that chips cannot be deployed until the data center shell is ready, a constraint that underpins his target of $230 billion in AI semiconductor revenue by fiscal 2028, as detailed in Broadcom's investor communications. The company guided Q4 2026 AI revenue to $21.7 billion. Both numbers assume a construction pipeline that keeps opening new square footage at the rate the last three years established.
The timeline mismatch is what makes Sacks' warning concrete rather than theoretical. A large hyperscale data center takes two to three years from permit application to rack-up. Fiscal 2028 for both Nvidia and Broadcom is roughly 15 to 18 months away. That means the permits for the data centers that need to absorb their current backlogs either already exist today or they are already running late. Sacks' argument is that the next generation of permits, the ones that would support demand from 2028 through 2030, are the ones at risk. If those permits dry up, the hardware manufactured to fill that capacity runs out of homes, creating exactly the scenario he once called "a disaster for everyone."
How State Governors Became the Chokepoint in a .3T Buildout Cycle
Hyperscaler capital spending reached nearly $800 billion in 2026. Analysts tracking the sector project it to hit $1.3 trillion in 2027, the majority already committed to building, expanding, or equipping data centers. That is a capital deployment comparable to the annual GDP of a mid-size European economy, poured into infrastructure that requires local approval at every material step: zoning, land use, grid interconnection agreements, water rights for cooling, and in many jurisdictions, environmental review.
Sacks identified governors of the largest data center states as the new decision-makers in this chain. Speaking on the All-In Podcast, he described them as "basically indicating that they favor a moratorium" until certain conditions are met. His base case is that fully permitted projects get built through 2027, but that few new approvals will come through for the years following. The All-In hosts noted that if approval barriers rise materially in the United States, project pipelines worth more than $50 billion could relocate to Canada, Mexico, Malaysia, or the Middle East, taking the associated jobs, tax revenue, and energy infrastructure with them.
The mechanism matters because data centers are not software that can be rerouted. Each facility represents years of local permitting, grid interconnection, and in some cases zoning variances that require sustained government support. Virginia alone accounts for more than a third of all US data center capacity, according to Data Center Knowledge. As MarketScale notes, rack power densities are now approaching 4,000 watts per chip, further concentrating the buildout around facilities already equipped for liquid cooling. Taken together, that makes a single state's permitting posture a material variable for the entire AI chip supply chain. When state-level political signals shift against data center development, the downstream effect is not a marginal delay: it is a termination of the permit queue for new sites, with no easy substitute geography inside US borders at scale.
Unlike traditional infrastructure projects, data centers are increasingly concentrated in clusters near existing grid capacity and fiber networks, further limiting relocation options. A hyperscaler cannot simply move a 500-megawatt campus to a friendlier state if the power and connectivity aren't available there. The result is that decisions made by a handful of state governors over the next 12 to 18 months will determine whether Sacks' 2028 timeline is a forecast or a fact.
Broadcom's $230B Bet Is Now Contingent on Existing Permits
Broadcom's AI revenue trajectory is among the most aggressive in semiconductors. The $230 billion fiscal 2028 target would make its AI segment larger than any chip company in history. That figure is not a vague aspiration: it has been stated by the CEO in the context of specific product roadmaps for custom AI accelerators built for hyperscaler customers who have themselves committed to multi-year procurement agreements. The near-term demonstration of trajectory, $21.7 billion in Q4 2026 AI revenue, is the quarterly stepping stone that investors are using to validate the multi-year extrapolation.
The dependency Broadcom's own CEO identified is the same one Sacks is warning about. If the data center shell isn't ready, the chips don't ship, and revenue doesn't recognize. Broadcom's fiscal 2028 target therefore maps directly onto the inventory of data centers that need to be permitted, grounded, powered, and fitted with cooling over the next 18 months. Any material slowdown in the number of new shells reaching completion narrows the window in which the target is achievable without delays or write-downs.
Arista Networks, which supplies the high-speed networking fabric interconnecting GPU clusters inside those data center shells, reported revenue climbing 58.5% year over year, with Q4 2026 guided to $3.9 billion to $4 billion. Its growth depends on the same buildout. Its customer base is narrow: a handful of hyperscalers generate the majority of its revenue, meaning Arista's forward guidance is structurally exposed to the deployment plans of operators who are themselves subject to the same permit risk. If those operators pause or slow new facility openings, Arista's order flow shows the signal first, well before Nvidia or Broadcom do.
Why Vertiv and Equinix Are Reading This Differently
Not everyone on the infrastructure stack reads the permit risk as bearish. Vertiv, which makes the power distribution and thermal management equipment that keeps data centers operating, reported organic order growth of more than 100% in its most recent quarter. Its CEO directly pushed back on the permit narrative, stating that power, permitting, and labor issues are "not slowing down growth." The company's customers typically order 12 to 18 months ahead of deployment, so Vertiv's current order book is predominantly for capacity that will open through 2027 and into early 2028, inside the window Sacks considers still safe. That timing distinction is crucial: Vertiv's strong orders reflect permit applications already in progress, not the new applications that may never get filed.
Equinix is positioned even more defensively. It has approximately 3 gigawatts of planned capacity with land, power, and permitting already largely secured. Its CEO made an argument that runs directly counter to the Sacks alarm: if other projects get cancelled or stalled, existing permitted capacity becomes scarcer and therefore more valuable, not less. That would strengthen Equinix's pricing power and leasing rates in exactly the scenario where chip companies face the most pressure.
Digital Realty reinforces that positioning. The company has 1.5 gigawatts under active construction, with 63% of that capacity already pre-leased, and a record $2.5 billion backlog providing revenue visibility into 2028. The pre-leasing rate indicates customers are committing to capacity before it opens, which signals that at least in the near term, demand is translating into binding contracts. The bear case for the operators, however, is straightforward: they benefit from permit scarcity only if AI development itself doesn't slow. If chip supply closes and foundation model training stalls because GPU clusters can't scale, Equinix's high-spec AI-ready capacity could face the same vacancy pressure as any specialized real estate built ahead of its demand curve. Critics also point out that Vertiv's CEO speaking from a backlog of existing orders is not the same as confirming that new orders are still arriving at the same pace.
The Dark GPU Scenario and What It Would Cost the Sector
The concept of "dark GPUs" has circulated in AI infrastructure discussions since at least 2025, when Sacks himself raised it as his primary risk concern, warning that "if you had dark GPUs, that'd be a disaster for everyone." At the time, he was worried about overbuild: too many GPUs manufactured for demand that didn't materialize quickly enough. His position has since reversed. Now he identifies the permit wall as the more immediate threat: too many GPUs manufactured for data centers that don't get built.
What makes the scenario financially severe is the structure of the semiconductor supply chain. Nvidia manufactures chips at TSMC under long-lead agreements that cannot easily be cancelled without penalties. TSMC's fabs run at near-full utilization and cannot dial back output without breaking commitments to other customers. The $279 billion in Nvidia's supply obligations represents hardware that will be manufactured regardless of what happens on the data center permit side. If the pipeline closes and those GPUs have no data center home, they face cancellation, indefinite delay, or discounting into alternative markets. None of those outcomes are painless given how constrained the broader chip supply chain remains.
The scale of the capex cycle also creates financial lock-in that limits how quickly any state moratorium could take effect. Hyperscaler commitments to specific data center sites, including land purchase, grid interconnection deposits, and equipment pre-orders, often run years ahead of the first rack being installed. Much of the 2027 buildout is already financed and in various stages of construction. What Sacks is warning about is the 2028-and-beyond wave: sites that are still in the permitting queue, have not yet broken ground, and represent the next leg of a buildout cycle that the semiconductor industry has already priced into multi-year revenue targets.
What to Watch Before the 2028 Decision Point
The leading indicator to track is state-level legislative action, not commentary. Sacks' warning is based on gubernatorial signals, not enacted rules. If any major data center state formalizes a moratorium or introduces new permitting standards that meaningfully extend approval timelines, the impact on semiconductor backlogs will be calculable: count the gigawatts in that state's permit queue, multiply by average server density, and project the chip delivery shortfall against the existing backlog. Virginia's outsized share of US capacity makes it the state to watch first, followed by Texas and the Pacific Northwest markets where Microsoft and Amazon have collectively committed more than $80 billion in infrastructure investment.
At the company level, Nvidia's next earnings call and whether its management acknowledges permit risk in its guidance for fiscal 2027 and 2028 will carry specific weight. Historically, Nvidia has described supply, not demand, as its limiting factor. If that framing shifts to reference permitting timelines, it would signal that the pipeline risk has moved from Sacks' podcast into the company's own planning. Broadcom's Q4 2026 actual revenue against its $21.7 billion AI guidance will indicate whether the near-term deployment pipeline is running as expected. A shortfall would not confirm the permit thesis immediately, but it would invite the question of which data center openings slipped past the quarter.
Vertiv's order trends remain the most forward-looking measurable signal. A company that books orders 12 to 18 months ahead of delivery will show permit slowdowns in its bookings well before construction starts slow. Equinix and Digital Realty's pre-leasing rates in Q1 2027 earnings would signal whether hyperscalers are committing to new capacity or pausing. A drop from Digital Realty's current 63% pre-leasing rate would indicate that demand commitments are softening. A rise in Equinix's rack pricing would confirm the scarcity dynamic its CEO described. Neither company reports Q1 2027 until early next year, leaving Sacks' 2028 timeline untested by hard data for at least another 90 days. The full company-by-company financial breakdown covering Nvidia, Broadcom, Arista, Vertiv, Equinix, and Digital Realty appears in 24/7 Wall St.
"By 2028, the pipeline will be closed or significantly reduced." David Sacks, All-In Podcast, November 7, 2026
Key Takeaways
- Nvidia holds $279 billion in outstanding supply obligations, all contingent on data center shells being physically permitted and completed to absorb the hardware
- Broadcom targets $230 billion in AI semiconductor revenue by fiscal 2028, a figure the company's own CEO has linked to the data center construction pipeline staying on pace
- State governors are signaling moratoriums on new data center permits, the specific chokepoint David Sacks identified on the November 7 All-In Podcast as threatening to close the pipeline by 2028
- Hyperscaler capex reached nearly $800 billion in 2026, with $1.3 trillion projected for 2027 and most already committed, creating financial lock-in that limits how fast any moratorium would bite
- Equinix and Digital Realty hold the defensive position, with 3 GW and 1.5 GW respectively of largely permitted capacity that becomes more valuable if new approvals dry up
Questions Worth Asking
- If state permit resistance intensifies, which hyperscalers have enough already-permitted capacity to meet their own AI compute targets through 2029, and which face the largest unhoused GPU risk?
- Would Nvidia and TSMC restructure their long-lead supply agreements if the US data center buildout rate dropped 30% from current projections, and what would that cost both companies?
- Does the concentration of AI data center capacity in a handful of US states represent a structural regulatory risk that the semiconductor industry's multi-hundred-billion-dollar backlog has systematically underpriced?
