AI data center buildout and community resistance

A Carnegie Endowment report published October 5 clarifies that Meta's 20-year Constellation Energy deal for the Clinton Clean Energy Center in Illinois buys clean energy credits rather than electricity for its data centers, with power from the plant continuing to flow to the regional grid.

A16z's markets report, published October 4, put 2026 AI infrastructure spending at $777 billion with $1.1 trillion forecast for 2027 and found only 2.2% of US households paid for AI services as of April; the buildout was also described as now exceeding all prior US infrastructure, including railroads and electrification, not only oil and gas.

Kansas City Fed President Jeff Schmid called on October 2 for regulators to assess whether AI's network of companies and contracts poses systemic financial risk comparable to too-big-to-fail banks, amid $765 billion in global AI capital expenditure in 2026.

McKinsey added that falling per-token prices do not lower total AI bills because agentic applications consume substantially more compute, with government costs moving from near-zero to material, and credit rating agency Egan-Jones called complete economic disruption from AI "all but certain".

Micron reports record revenue of $54.2 billion, up 379% year-over-year, with data center revenue up more than elevenfold to $18 billion

Bain projected on September 30 that AI must generate $6 trillion in annual revenue by 2031 to justify current data-center spending, triple its prior target, with existing consumer and enterprise services supplying only $1.8 trillion of that.

Separately, hyperscalers are issuing bonds at enough volume that investors now demand roughly 0.25 percentage points more yield on their debt, a shift Fed Chair Kevin Warsh has linked to AI-driven fundraising activity lifting Treasury yields.

Goldman Sachs Research on September 29 put 2026 global AI investment at $1.019 trillion, roughly $200 billion above hyperscaler capex figures, while Brookings put total AI infrastructure investment through 2032 at $10.3 trillion.

Bain, via Bloomberg, raised its AI revenue requirement to $6 trillion by 2031, three times its year-ago figure, with a $4.2 trillion gap existing services cannot close; Sequoia's David Cahn found a $600 billion annual shortfall between spending and ecosystem revenue, widening into 2026.

On September 27, company-level projections showed Google and Amazon expected to spend over $1 trillion combined on AI infrastructure, with Truist projecting AI cloud ARR reaching $2.1 trillion by 2030 as inference displaces training as the primary demand driver.

A September 28 analysis found debt-funded compute clusters on non-Nvidia chips likely cost more to finance than they save, because lenders charge about 2.5 percentage points more on GPU loans than on data-center facility loans of the same credit grade.

Kansas City Fed President Jeff Schmid calls on regulators to assess whether AI industry concentration poses systemic financial risk

Goldman Sachs on September 25-26 raised combined hyperscaler AI infrastructure projections to $1.2 trillion in 2027 and $1.4 trillion in 2028, about 66% above start-of-2026 estimates, noting a $250 billion miss or beat in 2027 spending would shift S&P 500 earnings growth by roughly 6 percentage points.

A Columbia Business School report found the buildout's required return of about $5.50 per GPU-hour at full utilization falls within the current $6-10+ market range, meaning the central 182.7-gigawatt scenario does not require compute prices to rise. Akamai signed a seven-year, $11.6 billion contract to run Anthropic compute workloads, with an option for $9 billion more tied to future spending milestones.

Moody's data published September 24 showed the five largest hyperscalers carry roughly $2.8 trillion in combined future obligations, up from about $350 billion in 2023, with over $1 trillion in uncommenced data-center leases, as global AI capital expenditure reached $765 billion in 2026, passing oil and gas industry capex.

Elon Musk separately disclosed that xAI's Colossus 2 targets up to 660,000 GB300 GPUs in phased deployments through late 2026.

Epoch AI research published September 23 found that the cost to reach equivalent benchmark performance has fallen roughly 47% per quarter since 2023, which Epoch describes as 54 times faster than electricity prices fell before 1973, 18 times faster than lithium batteries, and 6 times faster than compute.

Reaching a 75% score on GPQA Diamond fell from roughly $0.30 per question with OpenAI's o3 to roughly $0.0004 with GPT-5.6 Luna about 18 months later, a 725x reduction.

On September 22, SB Energy deferred its planned roughly $50 billion IPO after investors rejected its $178 billion capex program, reliance on OpenAI as tenant, and project-level debt, with NVIDIA's $105 billion guarantee limited to specified lease obligations at PORTS-Pike.

Credit data placed AI-related bonds at 18% of the investment-grade corporate bond index, and Oracle's 5-year credit default swap spreads have more than tripled since September 2025. Goldman Sachs estimated 2026 global AI investment above $1 trillion, and local opposition disrupted 45 US data center projects worth $68 billion in Q2.

DRAM prices reported up roughly 171% year-over-year as producers shift to HBM; Micron's AI memory reportedly sold out through end of 2026

TrendForce projects global data center power capacity to more than double to 211.1 GW by 2027; EPRI estimates US supply-demand gap could surpass 170 GW by 2030

On September 18, Crusoe closed a $3.9 billion Series F at a $30.9 billion valuation for AI factory expansion, Ayar Labs completed a $650 million Series E for optical interconnects, and Together AI introduced preemptible compute at half standard on-demand rates.

A SemiAnalysis investigation found Neocloud GPU capacity flows through layered reseller chains with sub-tenant workloads on hardware the immediate provider does not own, and the Commerce Department cited national security concerns in directing Kalshi to remove its AI-compute future curve, though a Commerce spokesman denied the report.

Emerald AI, Google, and NVIDIA launched the AI Energy Management Alliance on September 16, a standards body to let data centers act as controllable grid resources through workload shifting, storage discharge, and contingency response.

Lambda validated NVIDIA's DSX MaxLPS on Blackwell hardware, finding 19 nodes operating within a 16-node power budget raised cluster-wide token throughput 24% and improved performance per watt 23%. At the AI Infra Summit, NVIDIA announced the industry metric is shifting from peak FLOPS to validated agentic tokens per megawatt.

In a September 15 forecast, BloombergNEF projected US data centers will consume 20% of national power by 2035, modeling a 19 GW supply shortfall even after record grid additions, with at least 15 coal plants postponing retirements since January 2025 to meet AI demand.

Emerald AI and NVIDIA demonstrated a flexible-load program with Silicon Valley Power that throttles low-priority AI workloads on real-time grid signals while keeping critical jobs running, adding a demand-side management dimension to the energy constraint picture. Qualcomm and Amazon also announced a multi-generation agreement to co-develop custom AI inference chips and high-speed connectivity hardware for AWS data centers.

Google announced a $15.1 billion investment in Finland's AI infrastructure, its largest in Europe, as McKinsey projected the US will need 130,000 additional electricians and 240,000 construction laborers by 2030 for AI infrastructure, with experienced data center electricians in high-demand markets earning up to $280,000.

H100 median marketplace rates have fallen to $3.38/hr from highs near $7-8/hr in 2024 while B200 chips remain scarce at $5.63-$8.64/hr, and Together AI introduced preemptible compute at half the on-demand rate. Investor Gavin Baker argued memory makers at 3-5x PE and Nvidia at a comparatively low PE cannot both reflect the same spending wave, and that profit pools concentrate in whatever segment is hardest to substitute; Jason Calacanis separately argued Nvidia benefits regardless of which AI company wins.

Satya Nadella disclosed Microsoft is repurposing older fleet GPUs to accelerate Microsoft Fabric, achieving more than 7x performance gains

NVIDIA announced partnership with Australian data-center providers targeting up to 2 GW of AI infrastructure capacity by 2027

On September 11, Sam Altman, at YC Startup School 2026, projected per-capita token usage growing from roughly 100,000 per month today to 500 billion within six and a half years, while Eric Schmidt argued financial capacity is the binding constraint on the buildout, estimating roughly $50 billion per gigawatt.

Supply bottlenecks also sharpened: indium phosphide demand for optical transceivers is running roughly three to four times available wafer capacity, prompting Nvidia to prepay billions to lock in supply, and Zuckerberg backed a $115 million electrician-training program to address a structural workforce shortage.

Nomura modeled global memory revenue at $3.675 trillion by 2030, roughly 20x from 2024 levels, with AI data center memory alone comprising nearly half that total.

The same September 9, Andreessen Horowitz launched a $1.1 billion AI hardware infrastructure fund and SK Hynix broke ground on a $4 billion HBM packaging facility in Indiana backed by the CHIPS Act. At the G20, OpenAI CEO Sam Altman, appearing alongside US Commerce Secretary Howard Lutnick, compared rejecting AI to a country rejecting electricity.

Gartner's August 2026 forecast put memory revenue at $837 billion for 2026 and over $1 trillion for 2027, tracking to 54% of semiconductor revenue; US power demand needs 36.3 GW of data center additions by 2027 against 8.5 GW added in 2025.

At the G20 on September 9, Nvidia CEO Jensen Huang argued AI investment could add $20 to $50 trillion in global economic benefit, and Dylan Patel predicted an ASML EUV bottleneck will limit AI compute by decade's end.

Token consumption data published September 8 documents the demand driving the AI capex buildout: Google's processing grew from 9.7 trillion tokens per month in May 2024 to over 3.2 quadrillion by May 2026, and agents account for 71% of all tokens on OpenRouter, growing 14x faster than human usage.

Goldman Sachs expects 24x token growth between 2026 and 2030, and Evercore projects roughly 250 gigawatts of new data center capacity needed to keep pace.

ByteDance secured a $29.6 billion three-year loan, portion expected for overseas AI and data-center expansion

UBS forecasts HBM demand growing from 32.9 billion gigabits in 2026 to 58 billion gigabits in 2027, with Nvidia's share falling from 58% to 41%

On September 7, Broadcom reported AI semiconductor revenue of $16.7 billion last quarter, up 221% year-over-year, and Federal Reserve Chairman Warsh at Jackson Hole said token sales for the two leading AI labs exceed $100 billion annually, up more than 500%.

Data Center Watch tallied $130 billion in blocked data center projects with more than 800 opposition groups and at least 37 arrests in 2026; the White House introduced a Ratepayer Protection Pledge and a $50 million campaign linked to Greg Brockman launched to counter the backlash.

Lambda CTO Balaban argues for AI compute as national strategic asset, citing local tax revenue impacts of data centers

Dylan Patel argues OpenAI and Anthropic are becoming the largest compute infrastructure buyers, with Anthropic's reported $45B compute agreement cited as evidence.

Carnegie Endowment reports that Meta's Constellation Energy deal involves clean energy credits, not direct electricity purchase for data centers

OpenAI and Oracle abandon Texas Stargate expansion; broader Oracle-OpenAI contract valued at $300B lifetime remains on track; Oracle announces $50B debt and equity raise

Meta and Microsoft each sign 20-year power purchase agreements with Constellation Energy for nuclear power