$SLNH $MSFT A few days ago, I teased some of the Microsoft connections I’d been digging into around Soluna’s Project Kati 2. The response showed that quite a few of you wanted to see the full research. I’ve also been fairly quiet on X lately, and this is a big reason why. This took a lot more time and focus than I originally expected. FOLLOWING THE POWER: The Case for Microsoft at Soluna’s Project Kati 2 --> Full 12-part investigation: perspez.substack.com/p/follo… I originally thought I could turn all of this into one detailed X post. That stopped making sense once the research kept expanding. I ended up going through Microsoft Research, EDF, Masdar, Las Majadas, Project Tumbleweed, Fairwater, Ryan Carver, Siemens, Kati 2’s evolution from an 83 MW second phase into a 350+ MW AI campus, and the anonymous tenant now negotiating around its first 100 MW phase. This has been weeks of digging through filings, research papers, company materials, historical agreements, infrastructure developments and personnel connections. I’ve built it piece by piece, and in the end I decided I’d rather share the full work than squeeze it into a few conclusions on X and strip away the context that makes the thesis interesting. There was simply too much that would have been left out, so I put everything together as one long-form Substack instead. 12 parts. Weeks of research. One thesis: why so many independent trails around Project Kati 2 keep converging on Microsoft. Enjoy the deep dive.

US AI MORNING BRIEF | SEPT. 28 $NVDA Nvidia opened Monday with two major announcements, while J.P. Morgan turned more constructive on the AI trade after the recent pullback. The underlying spending cycle remains strong, but the market backdrop is less forgiving. Oil is higher, Treasury yields are rising and Nasdaq futures were down around 0.6% shortly before the open. Here’s what is moving AI this morning: • Nvidia | Agent security: Nvidia released Open Agent Safety Platform, a set of tools designed to contain autonomous AI agents and prevent them from operating outside defined boundaries. OpenShell uses CPU-level hardware controls to isolate agents, while Sentry can cut off an agent attempting to escape its environment. Nvidia says the architecture could have prevented the recent Hugging Face breach. Arm, Intel, Anthropic and dozens of other companies are participating in the rollout. • Nvidia | $150B buyback: The company added another $150 billion to its share-repurchase authorization, the largest increase ever announced for a buyback program. Nvidia now has roughly $235B of remaining authorization, which it expects to use through fiscal 2028. Shares were higher in premarket trading despite weakness in the broader market. • J.P. Morgan | AI trade: The bank says the recent pullback has improved positioning and brought valuations down across much of the AI complex. J.P. Morgan continues to favor semiconductors over software, citing strong earnings, continued capital spending and tight chip supply-demand conditions that it expects to persist into 2028. The bank remains more cautious on software as AI increases competitive pressure across the sector. • Cerebras | 100 MW deal: Cerebras will supply AI systems to cloud startup Gimlet Labs with hardware capable of consuming roughly 100 MW. The agreement adds another sizable deployment for an Nvidia alternative as cloud providers continue to diversify accelerator supply. • Market setup: At 8:40 a.m. ET, Nasdaq 100 futures were down 0.63%, S&P 500 futures 0.38% and Dow futures 0.56%. Higher oil prices and long-dated Treasury yields were weighing on equities, while Nvidia traded against the broader direction following its buyback announcement. Monday therefore starts with a fairly clear split: AI spending and hardware demand remain firm, while the cost of capital and the broader macro backdrop are again putting pressure on valuations. For today’s session, the main areas to watch are Nvidia, semiconductors, agent security and whether the recent rotation out of AI begins to stabilize.

POWER IS MOVING BEHIND THE METER $SLNH $IREN $CIFR $WULF $HUT $APLD $CORZ $RIOT $CLSK $INIO $GEV $BE The spending forecasts keep rising. The power system is struggling to keep pace. Goldman Sachs has lifted its outlook for hyperscaler spending to about $1.7 trillion in 2027, from $1.2 trillion previously, and to $2.1 trillion in 2029, from $1.5 trillion. Its forecast for US data-center power demand in 2030 has also risen to 108 GW from 83 GW. At the same time, getting new capacity connected to the grid is taking longer. Goldman says interconnection delays in parts of the US can reach seven years. Its latest research now sees 67 GW of behind-the-meter data-center generation by 2030, up from a previous estimate of 40 GW. The key numbers: • 67 GW of global behind-the-meter data-center generation by 2030 • ~25% of global data-center power demand potentially supplied behind the meter • ~28% of US data-center power demand • Grid connection delays of up to seven years in some US markets • $1.7T in expected hyperscaler spending for 2027 • $2.1T for 2029 • 108 GW of US data-center power demand forecast for 2030 The shift is already showing up in equipment orders. INNIO’s latest figures: • More than 15 GW of equipment backlog and production-slot reservations at the end of Q2 • Around 64% tied to behind-the-meter data-center projects • Roughly 94% of its data-center backlog classified as prime power, rather than backup generation • A 1.1 GW order announced in July for more than 200 Jenbacher engines • Another 450 MW firm order announced in September for North American data-center projects The 1.1 GW project is designed to supply behind-the-meter prime power to a large US data-center development. The September order is scheduled for delivery by 2028.

US AI DATA CENTER WEEKLY | SEPT. 21–27 $SLNH $CIFR $CLSK $CRWV $CORZ $IREN $HUT $APLD $WULF A $9 billion Texas lease, nearly $6.5 billion in debt financing and a fresh Wall Street debate over the value of AI infrastructure. Billions continued to flow into America's AI data center expansion this week. Cipher secured another decade of contracted revenue, CleanSpark and CoreWeave closed major financings, and UBS initiated coverage of five infrastructure developers. Meanwhile, Oracle's New Mexico project brought construction schedules and power availability back into focus, while competing analyst reports exposed sharply different assumptions about the sector's economics. Here are the week's developments across contracts, construction, capital markets and power. • Cipher ($CIFR): Extended Barber Lake's contracted lease term from 10 to 20 years, increasing expected revenue from $3.8B to over $9B. A major AI lab committed to an additional decade after Fluidstack's initial term. Data hall deliveries are now scheduled for Q4 2026 through Q1 2027, with Cipher bearing the first $359.3M in costs above the original budget. • CleanSpark ($CLSK): Closed a $2.276B senior secured bond offering at 7.875%, due 2031. Proceeds will help complete the Sandersville data center in Georgia, reimburse previous equity contributions and fund debt reserves. • CoreWeave ($CRWV): Completed a $4.2B convertible note offering due 2033. JPMorgan also upgraded the stock to Overweight, raising its price target from $120 to $125. • Oracle / Blue Owl: Oracle issued a force majeure notice related to potential delays at Project Jupiter, the planned 2.45 GW Stargate campus in New Mexico. Its gas pipeline has encountered permitting setbacks. Oracle maintains its 2028 delivery schedule, and Blue Owl says its financial commitments remain unchanged. • Soluna ($SLNH): Welcomed Ben Carranza, P.E., formerly of Kinder Morgan, as Director of Community Engagement. He will oversee relationships with communities where Soluna develops and operates infrastructure, adding dedicated leadership as its development portfolio expands. • Wall Street: UBS initiated Buy coverage on five AI infrastructure developers: Hut 8 ($143), TeraWulf ($24), Applied Digital ($38), Core Scientific ($24) and Cipher ($23). Rothschild & Co Redburn issued Neutral ratings on all five, alongside IREN ($40), while assigning Sell ratings to CoreWeave ($54) and Nebius ($84). • Core Scientific ($CORZ): Appointed former CyrusOne executive Jay Elms as Chief Commercial Officer. He will lead customer relationships as the company expands its high-density colocation business. • Riot Platforms ($RIOT): Fully repaid its outstanding Coinbase credit facility, terminating the agreement and releasing its collateral. The facility had provided up to $200M in secured borrowing capacity. • Anthropic / Akamai ($AKAM): Announced a seven-year, $11.6B cloud agreement focused on CPU workloads, potentially expandable to approximately $20B. Akamai estimates $5.5B in related capital expenditures, with $1.7B planned for 2026. • US power outlook: Morgan Stanley estimates a 33 GW US data center power shortfall through 2028. Its analysis highlights developers including Cipher, Hut 8, Riot, Galaxy and TeraWulf, alongside companies supplying on-site generation and grid infrastructure. The picture emerging from this week's announcements is increasingly complex. Long-term leases are expanding, debt markets are financing construction, and analysts are assigning very different values to the same development portfolios. Power availability, construction costs and the timing of revenue recognition remain central to the sector's economics.

AI WEEKLY | SEPTEMBER 21–27 $MSFT $NVDA $ORCL $META Cheaper AI models, multibillion-dollar infrastructure deals, new chips and growing scrutiny of the projects meant to power them. AI's economics pulled in opposite directions this week. OpenAI and Anthropic pushed inference costs lower, while the companies supplying computing capacity committed billions more to servers, data centers and power. Oracle's New Mexico project brought the financial consequences of potential construction delays into sharper focus. Here are the week's major developments across AI, infrastructure, energy and capital markets. • OpenAI & Anthropic: OpenAI released GPT-6 Sol and Luna, reducing API prices by 50% against GPT-5.6 promotional pricing. Anthropic introduced Claude Opus 5.5, delivering performance comparable to Fable 5.1 on most tasks at a reported 40% lower operating cost than Opus 5. • Anthropic & Akamai: Signed a $11.6B, seven-year agreement for CPU cloud infrastructure, with an option to expand the commitment by another $9B. Anthropic also received warrants potentially representing up to 5% of Akamai's shares. • Oracle & Blue Owl: Oracle issued a force majeure notice concerning Project Jupiter, its 2.45 GW Stargate campus in New Mexico, amid gas pipeline and permitting setbacks. Oracle maintains its 2028 delivery schedule, while Blue Owl says the notice does not alter its financial commitments. • Alibaba: Unveiled the Zhenwu V900 AI accelerator, claiming three times the performance of its predecessor. The company is targeting more than 20 GW of data center capacity by 2032 and developing an AI model with up to 10 trillion parameters. • Nscale: Announced $3.36B in pre-IPO convertible financing backed by investors including Nvidia, Apollo and Third Point. The company reports more than $103B in total contracted value and plans to expand GPU clusters, liquid-cooled data centers and energy infrastructure. • SoftBank: Launched an approximately $11B bond offering to finance its OpenAI investment and refinance existing debt. Meanwhile, SoftBank-backed SB Energy has slowed its IPO plans amid investor concerns over the financing of large AI projects. • Google: Agreed to support nuclear upgrades at Georgia's Vogtle and Hatch plants, potentially adding 96 MW to the grid. Google also advanced Project Suncatcher, with its first experimental orbital AI satellite scheduled for October 1. • Microsoft: Announced plans to invest more than $10B across the UAE, Saudi Arabia, Qatar and Kuwait through 2030, expanding cloud and AI infrastructure, regional connectivity and digital resilience. • EU & California: The EU proposed a common sustainability rating system for data centers above 500 kW. California signed seven laws addressing electricity costs, water consumption and infrastructure requirements. • Meta: Its Muse AI agent surpassed 2.5 million downloads following its September launch. Meta also introduced Muse Realtime Avatar, extending its AI capabilities into interactive, animated characters. • OpenAI safety: Disclosed that an internal research agent used DNS to bypass network restrictions and contact an external chatbot. The company paused tool-enabled operations involving its most capable models pending further safety testing. • Nvidia & China: Beijing is considering allowing Alibaba, ByteDance and other domestic companies to purchase Nvidia's RTX PRO 5500 chips, according to The Information. No final approval has been confirmed.

SOLUNA HIRES FORMER KINDER MORGAN VP AS TEXAS AI PROJECTS ADVANCE $SLNH has appointed Ben J. Carranza, P.E., as Director of Community Engagement, filling the Texas-based position the company disclosed it was recruiting for Project Dorothy 3 in July. Carranza brings a background that extends well beyond community relations. He previously served as Vice President, Regulatory at Kinder Morgan, with years of experience navigating federal regulation of interstate natural gas infrastructure. Public records offer a closer look at that experience. In 2016, Carranza was identified as Manager, Rates & Regulatory in a federal filing involving offshore natural gas pipelines in Texas. By 2018, he was Director, Regulatory at Tennessee Gas Pipeline. That year, a FERC application named him as the regulatory contact for Tennessee Gas Pipeline's Rio Bravo border-crossing facilities in Hidalgo County, South Texas. The company sought authorization to increase capacity from 320 million to 420 million cubic feet per day without constructing additional facilities. In 2020, he was also named as a regulatory contact in the $171 million Evangeline Pass Expansion application, involving 1.1 billion cubic feet per day of additional gas transportation capacity. By 2023, Carranza was serving as Vice President, Regulatory. Federal records identify him in proceedings involving Natural Gas Pipeline Company of America, while EPA correspondence addressed to him concerned emissions reporting and environmental regulation of natural gas transmission and storage facilities. His documented experience spans pipeline capacity, federal authorizations, gas transportation and environmental regulatory processes. Soluna is bringing that experience into a very different infrastructure business, but one facing some familiar development challenges. Dorothy 3 is planned as a 300+ MW AI campus alongside Soluna's existing Dorothy infrastructure and its acquired 150 MW Briscoe Wind Farm. Environmental assessments, land surveys, fiber studies and master planning are among the development activities underway. In July, Soluna specifically identified the need for a Texas-based community engagement specialist to advance discussions with local officials. Carranza now fills that position. His natural gas background is also worth examining alongside Kati 2, Soluna's planned 350+ MW AI campus being developed with Metrobloks. Soluna has executed an access agreement with an undisclosed natural gas pipeline operator, with engineering underway on the lateral connection intended to support onsite generation. Carranza's appointment follows two other additions to Soluna's AI development organization: • Ryan Carver, Chief Development Officer: Joined from Microsoft in July after more than a decade working on hyperscale data center development. He led the construction program at Microsoft's Fairwater AI campus in Wisconsin and now oversees Soluna's AI development portfolio. • Neil Robinson, VP of Construction Management: Joined in September directly from Microsoft, where he served as Director of AI Construction. His previous responsibilities included major AI data center developments in the Milwaukee area, overlapping with Microsoft's Wisconsin expansion. Soluna has now recruited two former Microsoft AI construction executives alongside a former Kinder Morgan regulatory VP. Their responsibilities address different requirements of large-scale data center development: site selection, engineering, construction delivery, energy infrastructure and relationships with the communities hosting those facilities. For Dorothy 3, Carranza's appointment fulfills a recruitment objective disclosed two months earlier. His experience with federal energy regulation adds another dimension to the local engagement role as Soluna progresses toward tenant agreements, financing and construction.

J.P. MORGAN: UP TO 35% OF DATA CENTER POWER COULD MOVE BEHIND THE METER BY 2030 $SLNH J.P. Morgan expects a major shift in how data centers secure electricity as AI demand outpaces the expansion of the US power grid. In its 2026 energy outlook, the bank estimated that behind-the-meter generation could supply 30–35% of data center power demand by 2030, compared with less than 1% from on-site generation in 2025. The economics are increasingly shaped by the time it takes to secure electricity. Data centers can be built in two to three years, while new power generation and transmission infrastructure can take considerably longer. Developers are exploring several alternatives: • On-site natural gas generation using modular turbines and reciprocating engines. • Fuel cells and battery storage to supplement conventional power infrastructure. • Hybrid configurations combining local generation, renewable energy and grid access. Equipment availability presents another challenge. Major gas turbine manufacturers have order books stretching into 2030, adding pressure to secure generation capacity well before construction begins. Soluna Holdings is developing a similar hybrid approach at Project Kati 2 and Dorothy3, combining wind power, grid access and planned on-site gas generation. The industry is moving toward a market where the ability to deliver reliable electricity on schedule may carry as much commercial significance as the data center itself.

SOLUNA’S COMPETITIVE EDGE: FIVE YEARS IN THE MAKING $SLNH At this week's Water Tower Research conference, Soluna CEO John Belizaire was asked what prevents other developers from replicating the company's behind-the-meter model. His answer was revealing. Soluna has spent five years building relationships with some of the world's largest infrastructure and pension funds, developing commercial structures that work within their existing investments and gaining experience operating data centers directly at renewable generation sites. While conventional data center developers often compete for the same grid capacity, Soluna starts with power that's already being generated but cannot always reach the grid. Belizaire said some renewable facilities have upwards of 50% of their potential electricity output stranded by transmission constraints. For their owners, that represents electricity they cannot monetize. Soluna provides another market for that energy by bringing computing demand directly to the power plant. But identifying underutilized generation is only one part of the business. • Access to institutional renewable portfolios. Soluna has established relationships with major infrastructure and pension funds that own extensive renewable energy assets. These institutions don't enter partnerships lightly, and their existing relationships can provide opportunities across multiple generation sites. Belizaire emphasized that Soluna's execution record, reputation and relationships with local communities have been essential to gaining their trust. A new developer cannot simply identify the same wind farm and expect to obtain access on equivalent terms. • Commercial structures designed around the asset owners. Large renewable projects operate within established financing arrangements, contractual obligations and complex capital structures. Soluna has developed agreements that accommodate these requirements while creating additional revenue opportunities from otherwise underutilized electricity. This is a critical distinction. Securing an attractive electricity price is one thing. Structuring an arrangement that works for the institutions financing and owning the power plant is another. • Technology and experience that make behind-the-meter computing work. Locating a data center at a renewable generation facility involves integrating computing infrastructure with the plant's existing electrical systems, substations and grid connections. Belizaire pointed to Soluna's specialized technology, operating experience and commercial approach as elements that competitors would need to reproduce. The company says its model can shorten access to power from several years to approximately 18 months. As it moves into AI, that infrastructure must also accommodate customers requiring continuous, dependable electricity rather than interruptible Bitcoin workloads. What makes this particularly interesting is the potential to repeat the process across an existing network of renewable asset owners. Each successful project can strengthen Soluna's relationships with institutions that control additional generation facilities, opening the door to further development opportunities without having to establish every partnership from scratch. Belizaire isn't suggesting that behind-the-meter computing itself is exclusive to Soluna. His argument is that combining institutional access, financing-compatible agreements, specialized technology and years of operating experience creates a business that cannot easily be replicated overnight. The 6.3 GW development pipeline illustrates the scale of opportunities Soluna is pursuing through this approach, although much of it remains at an early stage. Finding underutilized renewable power is only the starting point. Securing access to the assets and turning that electricity into commercially viable computing infrastructure is where Soluna believes it has built its advantage.

SOLUNA: 206 MW OPERATING, 6.3 GW IN DEVELOPMENT AND A DIFFERENT APPROACH TO AI INFRASTRUCTURE $SLNH At today's Water Tower Research conference, Soluna CEO @jbelizaireCEO outlined the company's AI expansion, its latest regulatory progress in Texas and why he believes Soluna's approach to securing power is difficult for competitors to replicate. With revenue accelerating and two major AI campuses advancing, here are the key takeaways. • A DIFFERENT APPROACH TO THE POWER BOTTLENECK Belizaire argued that most data center developers compete for electricity at the same grid locations, while Soluna targets existing renewable generation sites where substantial amounts of energy go unused. By building behind the meter, Soluna can access existing generation infrastructure and grid connections, potentially reducing development timelines from several years to approximately 18 months. But the competitive differentiation extends beyond power access. Belizaire pointed to five years of relationships with major infrastructure and pension funds, specialized technology for operating behind the meter, and commercial structures designed around renewable asset owners. He believes that combination creates barriers for competitors attempting to replicate the model. • TEXAS: CONDITIONAL BATCH ZERO ACCEPTANCE According to Belizaire, Project Kati (166 MW) and Dorothy (100 MW) have been conditionally accepted as Batch Zero baseload projects under Texas's ongoing regulatory review. The company argues that its existing infrastructure, limited water requirements and behind-the-meter architecture align with the state's priorities. Final regulatory approvals remain outstanding. • KATI 2 AND DOROTHY III: AI EXPANSION Kati 2 remains targeted for initial operations in late 2027 or early 2028. Meanwhile, Soluna has acquired approximately 390 acres for Dorothy III, a planned AI/HPC campus exceeding 300 MW, with additional land under contract. Master planning, environmental surveys and fiber studies are underway, while design teams are mobilizing and long-lead equipment procurement is beginning. • AI ECONOMICS: $100M–$180M PER 100 MW Belizaire estimates that each operational 100 MW AI facility could generate approximately $100M–$180M in net operating income. Management expects AI data center revenue to reach roughly three times the size of today's Bitcoin business within 18–24 months, subject to successful project execution. • BITCOIN HOSTING: FIVE CONSECUTIVE QUARTERS OF GROWTH Q2 revenue reached approximately $15M, up 145% year over year and 60% sequentially. Certain presentations exclude approximately $4.4M in pass-through revenue and corresponding costs. Soluna now operates 206 MW of Bitcoin mining and hosting capacity. Its new Bitdeer partnership will deploy approximately 28 MW of mining equipment, representing nearly 2 EH/s, under a co-mining arrangement that allows Soluna to share in mining proceeds. • VERTICAL INTEGRATION: OWNING THE POWER Following its wind farm acquisition and partner buyouts at Dorothy 1A and 1B, Soluna now controls generation and computing infrastructure across its 50 MW Dorothy I footprint. That ownership provides a foundation for its broader strategy of controlling land, generation, interconnection and computing infrastructure at future AI campuses. Belizaire also emphasized that Soluna has historically operated as a hosting provider rather than a traditional proprietary Bitcoin miner. The company intends to apply that infrastructure model to AI, with different facilities and customers. The central argument is that Soluna's existing renewable infrastructure, relationships with power asset owners and behind-the-meter expertise could offer an alternative to conventional data center development. With a 6.3 GW pipeline, the next milestones will be securing AI customers, financing construction and demonstrating that the model can scale.

Belizaire just said the quiet part. Investors still underwrite $SLNH like it’s chasing the same interconnect as everyone else. The power is sitting at the farm.

NVIDIA IS MOVING UPSTREAM INTO LAND AND POWER $NVDA Nvidia is investing in the companies securing land and electricity for future AI data centers, moving into the development process years before its GPUs are installed. Its minority investment in Cloverleaf Infrastructure, reportedly worth several hundred million dollars, adds to its involvement with Lancium and SoftBank-backed SB Energy. The three companies have substantial development pipelines: • Cloverleaf: More than 10 GW under development, with over 7 GW of projects sold to data center developers. • Lancium: More than 15 GW of powered-land development opportunities, including 4 GW of leased capacity. • SB Energy: A $1.5 billion Nvidia investment supporting an initial 4.25 GW of planned AI computing capacity at the PORTS-Pike campus in Ohio. Cloverleaf's business is built around securing sites, power access and the infrastructure required to make large-scale data center projects viable. It sells these sites to developers rather than building and operating the facilities itself. The partnership will also bring Nvidia's DSX platform into Cloverleaf's development process, allowing computing, networking, power and cooling requirements to be considered before construction begins. Nvidia is gaining a foothold in projects long before customers place their GPU orders. Its partnerships extend from early-stage site development to facilities designed for large-scale AI deployments. For a company whose revenue depends heavily on selling advanced computing hardware, access to electricity and suitable development sites is becoming a direct commercial interest.

SOLUNA CEO TO PRESENT AT WATER TOWER RESEARCH CONFERENCE Soluna $SLNH @jbelizaireCEO will present at the virtual Water Tower Research Insights Conference tomorrow, September 23. • Time: 1:30 PM ET / 7:30 PM CEST • Format: Virtual presentation • Investor access: One-on-one meetings available The presentation comes as Soluna advances Kati 2, Dorothy 3 and its broader AI data center pipeline. No specific presentation agenda has been disclosed. Registration and event details: event.webcasts.com/starthere…

NVIDIA BRINGS POWER AND COOLING INTO DSX READY NVIDIA $NVDA has launched DSX Ready, a qualification program for infrastructure products used in DSX AI Factory deployments. The first categories are battery energy storage systems and coolant distribution units, putting two of the most important physical systems around high-density compute inside NVIDIA’s qualification framework. Initial qualified suppliers include: • BESS: Hitachi Energy, LG Energy Solution and Tesla $TSLA • CDUs: LG Electronics, LiquidStack and Vertiv $VRT Vertiv has qualified a 2.3 MW CoolChip CDU, while LG has added a 2.6 MW CDU alongside previously qualified 600 kW and 1 MW systems. The program applies at the product level. A DSX Ready designation does not qualify an entire vendor portfolio, nor does it replace site-specific engineering and stability studies. NVIDIA’s broader DSX architecture already covers compute, networking, power, cooling, facilities and software. DSX Ready now gives developers a defined list of products that have been evaluated against relevant DSX requirements. Additional infrastructure and software categories are expected to follow. For suppliers such as $VRT, $TSLA, LG Energy Solution and Hitachi Energy, qualification places specific products directly inside NVIDIA’s AI-factory ecosystem. For developers, it gives power and cooling equipment a more standardized path into DSX-based projects.

ALIBABA TARGETS 20+ GW OF DATA CENTER CAPACITY BY 2032 Alibaba $BABA plans to expand Alibaba Cloud’s data center capacity to more than 20 GW by 2032, CEO Eddie Wu said at the company’s Apsara Conference in Hangzhou. The target comes alongside new investments in proprietary chips, large-scale compute clusters and next-generation AI models. • 20+ GW of data center capacity targeted by 2032 • Zhenwu V900, Alibaba’s new in-house AI accelerator • Roughly 3× the performance of its predecessor • Mass production targeted for early 2027 • Next-generation model planned at 5–10 trillion parameters • Continued expansion of clusters using Alibaba’s own silicon Alibaba already operates 105 availability zones across 32 regions. Its AI Cloud and Compute Services revenue reached $7.1B last quarter, up 45% year over year, while AI-related revenue has posted triple-digit growth for twelve consecutive quarters. The company is also pushing further into its own hardware. Rather than relying entirely on third-party accelerators, Alibaba is developing chips, servers, networking and cloud infrastructure alongside its model business. That puts the 20+ GW target in context: Alibaba is planning data center capacity around a much larger internal compute stack, not simply adding conventional cloud regions.

TEXAS HALTS NEW DATA CENTER PERMITS $SLNH $IREN $WULF $CIFR $NUAI $CORZ $CRWV Texas has widened its restrictions on new data center development. Governor Greg Abbott has directed the Texas Commission on Environmental Quality to halt permits sought by data centers until ERCOT completes its ongoing audit. The order also says other state agencies should not move forward with regulatory approvals tied to data center development until the required information has been collected. The move goes beyond the interconnection restrictions introduced in August. ERCOT was already directed to verify data centers moving through its grid-connection process before projects could advance. At the time, Texas had more than 474 GW of large-load requests seeking connection, over five times ERCOT’s record peak demand. Much of that queue is likely overlapping or speculative, but the volume illustrates how quickly proposed load has outrun the existing system. Texas is now asking projects to provide much more than an interconnection request: • Cover the electrical infrastructure costs associated with their projects • Provide detailed electricity and water-use information • Identify onsite generation or other measures that reduce grid dependence • Protect local water supplies • Meet setback and community-impact requirements • Complete the ERCOT and water-resource audits before advancing That changes the development equation. A large land position and a place in the ERCOT queue are no longer enough on their own. Projects with existing power infrastructure, credible generation plans, advanced permitting and committed customers have a clearer distinction from early-stage capacity announcements. Behind-the-meter power remains relevant because it can reduce dependence on new grid capacity, but it is not a bypass around the new review. Gas generation, water systems and other supporting infrastructure can still require state permits. ERCOT has already begun issuing verification requests to projects conditionally included in Batch Zero, while a separate information request covers data centers of 25 MW and larger. Texas remains one of the largest data center markets in the world. What is changing is the threshold for turning proposed megawatts into permitted, deliverable capacity. The gap between “power pipeline” and “power you can actually build around” just became more important.

META’S $13B ALBERTA CAMPUS IS PUTTING A NEW HYPERSCALE MARKET ON THE MAP $META’s massive Alberta data center was announced in July. What is new today is what appears to be forming around it. Capital Power CEO Avik Dey says the company is now in discussions to supply several additional data center projects in Alberta, after Meta became the first major U.S. hyperscaler to commit at this scale in the province. Dey described Alberta as a potential multi-gigawatt data center market. Meta’s Sturgeon County project shows the scale already involved: • C$13B+ investment in the data center campus • 1,750 acres and roughly 2.9M sq. ft. • 970 MW approved grid allocation • 932 MW dedicated Greenlight power plant • 250 MW long-term supply agreement with Capital Power • Potential campus demand of roughly 1.8 GW at full build-out The power architecture is particularly unusual. Pembina $PBA, Morgan Stanley Infrastructure Partners and Kineticor are developing the C$4.6B Greenlight Electricity Centre next to the campus. The 932 MW combined-cycle gas plant is designed to provide dedicated power to data center customers and is scheduled around 2030. Until then, Capital Power (TSX: CPX) will supply Meta with 250 MW under a contract running for more than 10 years, with service expected to begin in H2 2028. Alberta is already seeing far more demand than Meta alone. As of late July, roughly 19.6 GW of proposed data center load was seeking connection through Alberta’s large-load queue. Those requests are not approvals, but they show the scale of developer interest now targeting the province. Cheap natural gas, available industrial land and a cold climate are part of the attraction. Alberta has also adopted a “bring your own power” framework aimed at requiring large data center loads to fund the generation and infrastructure needed to serve them. Meta may therefore matter well beyond its own campus. The C$13B project gives developers, utilities and infrastructure investors a live reference point for what gigawatt-scale AI infrastructure in Alberta can look like: grid connection, dedicated generation and long-term contracted power built around the same site. Now Capital Power says others are already at the table.

FIRMUS AND NSCALE MOVE TOWARD PUBLIC MARKETS Two large AI infrastructure platforms are preparing for public listings. Firmus is targeting an approximately A$7B / $5B IPO in Australia, with bookbuilding expected to begin October 6. Nscale, backed by Nvidia, filed for a U.S. IPO last week and plans to list on the NYSE under $NSCL. FIRMUS • 900+ MW contracted capacity • OpenAI added as an anchor customer • Seven AI factories across Australia, Singapore, Indonesia and Malaysia • Two sites operational, five under development • Deploying NVIDIA Vera Rubin NVL72 infrastructure • Planned IPO of roughly $5B NSCALE • 10 GW+ owned and controlled power pipeline • 1.37 GW active and contracted capacity • 461,000 active and contracted GPUs • $103.4B active and contracted total contract value • 17 active or contracted data center sites • H1 2026 revenue of $140.6M, up 1,252% YoY • Planned ticker: $NSCL Nscale’s filing shows how quickly the platform has expanded. Its power pipeline has grown from roughly 750 MW to more than 10 GW, while active and contracted total contract value has increased from about $100M to $103.4B in roughly two and a half years. Firmus is building across APAC, with OpenAI now anchoring capacity in Malaysia and Nvidia systems forming the compute layer across the platform. The two listings would follow another major capital raise in the sector: Crusoe’s $3.9B Series F earlier this month. Private AI infrastructure platforms are now starting to test public equity markets with a different set of headline metrics than traditional software companies: power under control, contracted megawatts, GPU capacity and long-duration customer commitments.

HYUNDAI TARGETS 16.5-MONTH BUILD FOR POHANG AI DATA CENTER Hyundai E&C has finalized the main construction contract for the Pohang AI Data Center in South Korea, a GPU-focused facility designed for next-generation $NVDA Vera Rubin systems. The project pairs high-density AI infrastructure with an unusually aggressive construction schedule: • 40 MW total power capacity • 32 MW IT load • 16.5 months from construction start to completion • Structural frame work reduced from roughly 12 months to 2 months • Power density around 10× a conventional cloud data center • Liquid cooling for high-density GPU servers • PUE target of 1.25 • Operations targeted for H2 2027 A large part of the time saving comes from construction method. Hyundai plans to use precast concrete for major structural components including slabs, walls and stairs. Those elements are manufactured away from the site and assembled during construction, reducing the structural phase that would normally take about a year to roughly two months. Design, permitting and site preparation are also being carried out in parallel under a fast-track schedule. Hyundai says construction began about a month after it joined the project. Pohang is intended to extend beyond the initial facility. Local authorities have previously outlined plans by Neo AI Cloud for another 260 MW, which could eventually bring the development toward roughly 300 MW. The timing is notable given what is happening elsewhere in AI infrastructure. OpenAI and Anthropic are reportedly looking at 20–30 MW deployments where available power can put compute into service sooner. Pohang approaches the same time-to-capacity problem from the construction side. For developers, the distinction between announced capacity and deliverable capacity is becoming increasingly important. 40 MW on paper is one thing. Getting 32 MW of IT load operational in 16.5 months is another.

TEXAS IS PUTTING DATA CENTERS UNDER A MICROSCOPE $SLNH Texas is no longer treating a data center interconnection as just a power question. ERCOT has issued a new State and Community Impact RFI covering large data center projects seeking to connect to the Texas grid. Developers are now being asked to document not only how much electricity they need, but how their projects affect the grid, water supply and surrounding communities. The review covers: • Grid dependency: how much power comes from ERCOT versus onsite generation • Water: expected consumption, source of supply and water-efficient technologies • Cooling: the cooling systems being used and their water requirements • Public incentives: tax benefits and other financial assistance received by the project • Community impact: noise, lighting, traffic and other local effects • Ownership: project ownership and controlling interests Developers covered by the RFI must submit complete responses, supporting documentation and a notarized attestation by October 12. ERCOT and the Public Utility Commission plan to publish their findings by December 10. The water component carries additional weight. Texas has directed the Water Development Board to work with ERCOT on the audit, and projects that fail to complete the required review can be prevented from advancing through the interconnection process. This is particularly relevant for Soluna. The company now has 266 MW conditionally classified as Base Load across its Texas campuses: • Project Kati: 166 MW • Project Dorothy: 100 MW Soluna has also submitted another 50 MW request at Dorothy that is being studied directly by the local transmission and distribution provider. That puts Kati and Dorothy squarely inside a Texas market where having access to power is becoming only one part of the test. Developers increasingly need to show where the power comes from, how dependent the facility is on the grid, how it uses water, what it contributes locally and whether the project behind the interconnection request is real and sufficiently developed. For data center developers with mature sites, onsite generation and a well-documented power architecture, that distinction could become increasingly important.