COGNITO RIVER DELTA ♥ Support Follow on X
CognitoRiverDelta - AI News Evolved
Category rollup
Datacenters & cloud - 2026-W40

Week of September 28, 2026

This week marks a fundamental ontological shift in how the global industry views the physical layer of the digital economy. We are witnessing the transition of the data center from a "storage and retrieval" utility—a pas

AI News Evolved · cognitoriverdelta.ai

COGNITO RIVER DELTAAI News Evolved · https://cognitoriverdelta.ai
← All reports
Category rollup

Datacenters & cloud - 2026-W40

Week of September 28, 2026 · 20 min read Download PDF Share on X
Datacenters & cloudSpace-based compute & power

Datacenters & cloud · week 2026-W40: Sep 27 - Oct 03, 2026 · 1 subtopic(s) covered · 3022 words · expanded

Overview

This week marks a fundamental ontological shift in how the global industry views the physical layer of the digital economy. We are witnessing the transition of the data center from a "storage and retrieval" utility—a passive repository for bits—to a "superintelligence (SI) factory"—a term increasingly championed by industry leaders like Jensen Huang and Farzad Mesbahi. These facilities are being reimagined as active manufacturing plants that do not merely house data, but actively convert vast quantities of energy and compute into economic value through the production of intelligence. This rebranding is not merely a semantic exercise; it reflects a massive escalation in capital intensity and a total restructuring of the sector's risk profile. Projects are no longer measured in the hundreds of millions of dollars typical of enterprise cloud expansion, but in the hundreds of billions, representing a new epoch of "megaprojects." Consequently, the primary industry bottleneck has shifted from the availability of silicon to the fundamental, physical limits of the power grid, land availability, and the acquisition of local social license.

The developments of the week reveal a profound tension between two competing, high-stakes trajectories. On one side is a relentless, high-velocity expansion characterized by unprecedented infrastructure bets—such as NVIDIA backstopping a $105 billion project in Ohio and the emergence of $100 billion campuses in Kentucky—and a strategic pivot toward aggressive vertical integration. Companies like Tesla and SpaceX are moving to own the entire stack, from the semiconductor design to the energy source itself.

On the opposing side, a growing "friction of reality" is mounting, creating a significant drag on this momentum. This friction is manifesting as intense public and regulatory pushback, with recent polling indicating that 70% of Americans oppose data centers in their immediate neighborhoods. This social resistance is translating into massive project delays, with estimates suggesting that up to 50% of this year’s projected capacity is facing significant hurdles. Furthermore, increasing financial skepticism is creeping into the "AI trade," exemplified by SB Energy’s failed attempt to secure a $50 billion valuation despite lacking a single operating facility.

Ultimately, the story of this week is the "industrialization" of AI. The sector is moving out of its "software-first" era, defined by scalable code and light margins, and into a "physical-first" era, where success is dictated by the ability to navigate the complex, heavy, and often resistant landscape of terrestrial infrastructure: land acquisition, nuclear energy procurement, thermal management breakthroughs, and geopolitical maneuvering.

Earth-based data centers

The Rise of the "SI Factory" and the Megaproject Era

The scale of investment in terrestrial data centers has entered a new, almost unprecedented epoch of "megaprojects." We are no longer discussing individual server halls or even discrete data center campuses; we are seeing the emergence of massive industrial complexes that more closely resemble oil refineries, semiconductor fabrication plants, or heavy manufacturing hubs. This shift is epitomized by a handful of colossal capital commitments: NVIDIA is currently backstopping a $105 billion project in Ohio linked to OpenAI (to be operated by SB Energy), and a massive $100 billion proposal in Kentucky is being spearheaded by NextEra and Brookfield. In Canada, Meta is planning a C$13 billion facility in Alberta, while Tesla is contemplating the "TerraFab" in Texas, a semiconductor campus with a cost ceiling potentially reaching as high as $116 billion.

This shift is driven by a fundamental reassessment of what compute represents. Jensen Huang has emphasized that these "AI factories" are productive assets—industrial tools capable of generating rapid returns on capital by producing intelligence. Farzad Mesbahi reinforces this, noting that these facilities serve as primary drivers of macroeconomic growth and national GDP. However, this industrial scale brings a level of complexity and cost that the previous generation of data center operators is ill-equipped to handle. NVIDIA reports that these facilities are now being built on a "per megawatt" or "per gigawatt" basis, with costs reaching approximately $60 million per megawatt. This staggering CapEx requirement is necessitating a new class of corporate finance, as seen in NVIDIA’s recent move to partner with six investment firms to mobilize over $500 billion specifically for infrastructure.

The Energy-Compute Convergence

The most significant operational shift this week is the realization that data center developers are effectively transforming into energy companies. The "foundational layer" of the computing stack has moved from the chip to the electron. This has triggered a frantic, strategic race for firm, carbon-free baseload power, as the intermittent nature of solar and wind cannot meet the 99.999% availability requirements of hyperscale AI operations.

The emergence of Small Modular Reactors (SMRs) as a primary energy solution has become a dominant industry thread. Large-scale hyperscalers are no longer content with mere Power Purchase Agreements (PPAs); they are seeking direct involvement in nuclear deployment. Amazon has already partnered with X-energy to deploy 5GW of nuclear capacity by 2039, and Valar Atomics has proposed a massive 9.4GW SMR-powered campus in Utah, targeting its first live reactor as early as 2028. This pivot is driven by sheer necessity: Eric Schmidt has noted that AI data centers could account for up to 11% of total U.S. electricity demand by 2030. This scale of demand is forcing highly strategic, long-term energy arrangements, such as the 20-year agreement between Microsoft and Chevron for a Texas-based facility. The development model is evolving from "renting space" to "managing energy ecosystems," where developers take responsibility for the generation, storage, and distribution of the electricity required to run their own sites.

The Friction of Physicality: Regulation, Public Backlash, and Delays

As the physical footprint of these "SI factories" expands, they are hitting the hard limits of social and regulatory reality. A significant "NIMBY" (Not In My Backyard) movement is coalescing globally. Reports indicate that 70% of Americans oppose data centers in their immediate neighborhoods, a sentiment that spans the political spectrum—from environmentalists concerned about water consumption to populist movements concerned about land use and local aesthetics.

This social friction is not merely a PR problem; it is a massive operational bottleneck. Prof G Markets reports that 45 projects, totaling $68 billion, were blocked or delayed in a single three-month window, representing 30-50% of this year's projected capacity. Regulatory environments are tightening in response to this pressure. In Australia, Goodman was forced to withdraw a 90MW application in Sydney due to an "evolved" regulatory landscape. In Finland, grid connection reforms have effectively pushed data center projects down the priority queue behind other industrial needs. Even within the U.S., the data center debate has become a political flashpoint for midterm elections. In a strategic attempt to rebuild trust and address mounting suspicion, AWS has made the decision to drop non-disclosure agreements (NDAs) for its projects to foster greater transparency, even as its CEO, Matt Garman, warns that continued blocking of these projects could cause "irreparable harm" to national security and the economy.

Vertical Integration and the Strategic Arms Race

To bypass these mounting bottlenecks in power, land, and chips, the industry's largest players are pursuing aggressive vertical integration. We are witnessing a move to control the entire lifecycle of compute to ensure reliability and margin. Tesla's "TerraFab" project is a prime example, aiming to secure a proprietary supply of its upcoming "AI5" chips by controlling the entire chip lifecycle—from design and making to testing and packaging. Similarly, NVIDIA is moving to make its interconnect and rack infrastructure (via NVLink Fusion) essential to the physical layer, ensuring that even customers using competing accelerators remain tethered to NVIDIA’s ecosystem.

This integration is inextricably tied to the geopolitical struggle for technological supremacy. There is a growing consensus among U.S. leaders that owning AI infrastructure is a matter of national security. The U.S. and China are engaged in a strategic competition where compute and power capacity are viewed as critical tools of national power. In China, the state is actively consolidating data center construction into national hubs to create a network that can dynamically move computing tasks between regions. In the U.S., the "White House Accord on Super Intelligence" suggests a regulatory framework that will likely focus on controlling data center access and GPU allocation, effectively treating compute capacity as a critical utility for national defense.

Thermal and Resource Management Innovations

As power density skyrockets—with AI racks now requiring between 100kW and 132kW—traditional air cooling is becoming physically obsolete. This has spurred a rapid wave of innovation in thermal management and resource conservation. Companies like ZutaCore are bringing liquid cooling to high-stakes sectors like financial services, and Castrol has launched "Castrol CORE" to move from being a mere product provider to offering integrated thermal management.

Innovation is also being driven by the need to mitigate resource scarcity. MagPro has launched the AeroLev 1850, a "water-smart" cooling system designed to eliminate the need for evaporative cooling water in arid regions. Furthermore, the industry is exploring highly unconventional infrastructure to bypass terrestrial limits: NetworkOcean is developing floating solar-powered data centers in San Francisco Bay, and Xeal is exploring ways to utilize spare EV charging capacity to power edge AI inference. These innovations are direct responses to the fundamental tension between the massive thermal output of AI hardware and the increasing scarcity of water and stable electricity.

Cross-cutting themes

The overarching theme of this week is the collapse of the distinction between the digital and physical worlds. Historically, software and data were viewed as "weightless" assets that could scale infinitely with minimal friction. This week's developments prove the opposite: AI is incredibly "heavy." It requires massive amounts of physical land, billions of dollars in debt, thousands of megawatts of power, and vast quantities of water. The digital revolution has become an industrial revolution.

This creates a critical, ongoing tension between Scale and Scarcity. The industry is scaling at a "megaproject" level, throwing hundreds of billions of dollars at expansion, but it is doing so against a backdrop of increasing scarcity in the very resources required to sustain that scale: scarcity of power grid capacity, scarcity of water in arid regions, scarcity of social license, and scarcity of regulatory speed.

Finally, there is a profound shift from Software Economics to Industrial Economics. The value proposition of the sector is moving from the high-margin, low-CapEx world of software licenses to the capital-intensive, high-stakes world of heavy industry. This brings entirely new categories of risk that investors and operators must manage, including construction execution risk, the risk of severe financial overleveraging, and the potential for "circular financing" loops where the companies providing the capital are the same ones benefiting from the increased demand for their own hardware.

Where sources agree

  • Energy as the Foundational Bottleneck: There is near-unanimous consensus that energy is the primary limiting factor for AI scaling. Sources across the board describe energy not just as a utility, but as the "foundational layer" of the entire computing stack. Without a massive, stable supply of electrons, the entire AI stack remains theoretical.
  • The National Security Imperative: Sources agree that compute power and energy are now inextricable from national security. The ability to produce, house, and control intelligence domestically is viewed as a strategic necessity in the competition with China. This is evidenced by the calls from U.S. leadership to ensure AI intelligence is produced domestically.
  • Mounting Social and Regulatory Resistance: There is broad agreement that data centers are facing a global wave of political and social headwinds. This is evidenced by public opinion polls showing widespread opposition, the withdrawal of major applications (like Goodman in Australia), and the shifting regulatory focus toward the physical infrastructure itself.
  • The Rebranding to "SI Factories": There is a growing consensus among the most influential voices in the industry (Huang, Mesbahi) that the term "data center" is becoming obsolete. The shift toward "AI factory" or "superintelligence (SI) factory" is accepted as the correct way to describe the productive, rather than passive, nature of these facilities.

Where sources disagree

  • Economic Sustainability vs. The Overbuilding Bubble: A sharp divide exists between those who view the current build-out as an essential, high-return investment in national GDP (Jensen Huang, Farzad Mesbahi) and those who warn of a looming "overbuilding" bubble. Skeptics (Prof G Markets, Ed Zitron) argue that excessive CapEx could lead to a crash in lease rates and place unsustainable pressure on the bond market.
  • The Reality of Environmental Concerns: There is a significant disagreement regarding the severity of environmental impacts. Some analysts (The Economist) characterize public concerns about water usage and noise as "myths" or "misconceptions." Conversely, many other sources (CNBC Tech, Farzad Mesbahi) point to the massive political backlash and actual regulatory moratoriums as proof that these concerns are very real, politically potent, and capable of halting projects.
  • The Future of Compute Location: A radical tension exists between the current momentum of massive terrestrial investment (Meta in Canada, NextEra in Kentucky) and the long-term, highly speculative vision of analysts like Phil Bicil, who suggest that 99% of compute will eventually move away from Earth to space-based "compute fabrics."
  • The Necessity of Centralized Megaprojects: While the industry is moving toward gigawatt-scale campuses, some analysts (David Freeberg) argue that the vast majority of AI tasks do not actually require such massive, centralized facilities, suggesting a potential massive mismatch between the scale of current infrastructure build-out and actual workload requirements.

Numbers and claims to verify

  • $60 million per megawatt: The reported cost to build AI factories according to NVIDIA. (Needs verification against actual CapEx data from other hyperscalers).
  • 90% of global data center financing in the US: A claim made by Cathie Wood. (Requires verification against global FDI and infrastructure spending data).
  • 30x more water usage by golf facilities than data centers: A comparative claim made by Farzad Mesbahi. (Needs verification of water consumption metrics for both sectors).
  • 99.99% annual drop in AI inference costs: A claim made by Randy Kirk. (This represents an extreme rate of deflation and requires rigorous verification).
  • 30% reduction in Loudoun County property taxes: A claim made by The Economist regarding the impact of data center taxes. (Needs verification via local county tax records).

Investment and strategic implications

  • From Hardware to Infrastructure-as-a-Service: The investment thesis is fundamentally shifting. While the "chip trade" remains a core driver, the strategic imperative is moving toward the "physical layer." Investors should look toward energy production (specifically nuclear/SMRs), advanced thermal management (liquid cooling), and grid-edge technologies.
  • Vertical Integration as a Structural Moat: Companies that can control both the compute (chips/GPUs) and the power/infrastructure (through direct energy deals or on-site generation) will possess a significant competitive advantage. The moves by Tesla and NVIDIA suggest that the winners will be those who "own the stack" from the electron to the algorithm.
  • The Re-Rating of Data Center Developers: Data center developers are being re-rated as "power projects." Their enterprise value is increasingly tied to their ability to secure long-term, firm energy contracts and grid interconnects rather than just their ability to lease rack space or floor area.
  • Systemic Financial Risk and Overleveraging: The massive debt requirements for these megaprojects introduce new systemic risks to the broader economy. Investors must monitor for "circular financing" loops—where the same entities provide capital, hardware, and lease agreements—and the potential for a credit crunch if the projected returns on these "SI factories" do not materialize quickly enough to service the debt.

What to watch next week

  • SMR and Nuclear Deployment Milestones: Any regulatory or construction updates regarding X-energy or Valar Atomics projects will serve as critical bellwethers for the feasibility of the nuclear-data-center nexus.
  • Regulatory "Red Flags": Watch for any new "data center moratoriums" or significant policy shifts in major US or European hubs. Such news would signal that social and regulatory friction is successfully winning out over capital momentum.
  • NVIDIA Ecosystem Velocity: Look for further details on the $500 billion mobilization effort or updates on the deployment of "Reuben" and "Blackwell Ultra" models. This will indicate the velocity of the hardware refresh cycle and the health of the NVIDIA-led ecosystem.
  • The "Overbuilding" Signal: Monitor lease rates and occupancy data in primary hubs like Northern Virginia. Any softening in these markets would provide much-needed evidence for the "overbuilding" concerns raised by skeptics this week.

Appendix: Individual perspectives

  • Amy Webb: Dismisses apocalyptic water scarcity fears, suggesting data centers could actually provide climate solutions.
  • Brian Wang: Projects that mass-produced "mini hard" data center halls can scale to 10GW of capacity by late 2027 through a rapid, monthly completion cadence.
  • Cathie Wood: Argues that data center demand is a disinflationary force and that the US holds 90% of global data center financing.
  • David Carbutt: Suggests that high-performance, open-weight models allow 90% of AI tasks to run on local desktops, reducing the necessity for massive data centers.
  • Ed Zitron: Warns that the massive debt required for AI data centers is overleveraging the sector and straining the bond market.
  • Elon Musk: Sees terrestrial energy as the primary bottleneck and advocates for orbital compute and space-based solar power to bypass ground-based limits.
  • Eric Schmidt: Predicts AI data centers will account for up to 11% of total U.S. electricity demand by 2030.
  • Ezra Klein: Views AI infrastructure as a catalyst for modernizing the global energy grid (solar, nuclear, hydro).
  • Farzad Mesbahi: Rebrands data centers as "SI factories" that drive GDP; notes their water usage is lower than golf courses or almond orchards; emphasizes the geopolitical need for domestic construction.
  • Gwynne Shotwell: Identifies land price spikes and three-year electrical equipment lead times as major terrestrial constraints, making space-based compute a necessity.
  • Jason Calacanis: Predicts regulatory oversight will shift from AI models to physical infrastructure and GPU allocation within 12-18 months.
  • Jeff Lutz: Details terrestrial scaling bottlenecks including land acquisition, grid interconnect, and water cooling.
  • Jensen Huang: Views data centers as "AI factories" that transform energy and data into intelligence.
  • Jo Bhakdi: Contends that orbital AI data centers will achieve cost parity with terrestrial systems by 2029.
  • Steve Burke: Characterizes NVIDIA's strategy as "circular financing," where the company invests in the very entities that lease its hardware.

Sources

Related reports

Combined report

Transition to Physical and Orbital AI

How AI is crossing from software into the physical world and into orbit - robots, factories, energy and space-based compute, read together.

September 25, 2026 · Space-based compute & power · Datacenters & cloud
Daily AI Briefing

Multi-Trillion-Dollar Capital Expenditure Expansion

The global technology and investment landscape is undergoing a structural pivot toward physical AI, autonomous commercial transport, and massive orbital compute infrastructure, driven by multi-trillion-dollar capital expenditure plans across Elon Musk's corporate ecosystem. Industry analyses from le

September 26, 2026 · Space-based compute & power · Autonomous vehicles
Category rollup

Frontier labs & models - 2026-W40

The defining storyline of this week across the frontier model landscape is the definitive pivot from "AI as a chatbot" to "AI as a physical, agentic ecosystem." We are witnessing the end of the era where intelligence was

Week of September 28, 2026 · Frontier labs & models · Space-based compute & power

Informational analysis synthesized by AI from sourced, dated material, curated by a human. Treat specific claims as unverified until checked. Not financial advice.

About · How this is made · Corrections