SB Energy files for a Nasdaq IPO, targeting a $5-7 billion raise at a roughly $50 billion valuation

SoftBank-backed AI data center and energy infrastructure developer SB Energy has filed for an initial public offering on the Nasdaq Global Select Market and Nasdaq Texas, DCD reported on September 3. Reports place the target raise at 5 to 7 billion dollars, which would value the company at about 50 billion. J.P. Morgan, Goldman Sachs, Morgan Stanley, Citigroup and Mizuho are acting as joint lead book-running managers, and the company is also proposing a public offer of shares to retail investors who are UK tax residents. The prospectus is unusually candid about concentration risk, stating that SB Energy is "substantially dependent" on OpenAI as both tenant and equity investor, with SoftBank and OpenAI its only two customers. SB Energy is developing an up-to-10 gigawatt site in Pike County, Ohio, whose full capacity is slated to be leased to OpenAI, and it has been signed on to build and operate the 1.2 gigawatt Stargate data center in Milam County, Texas. Founded as a SoftBank subsidiary in 2011, the company also plans, alongside SoftBank, to invest at least 4.2 billion dollars in new regional grid infrastructure through a partnership with AEP Ohio. Two implications for the field. A public listing means quarterly disclosure on capital spend, schedule and tenant status for two of the largest programs in the country, which is free intelligence for anyone selling into them. It also means the financing for those programs now has to survive public market scrutiny rather than private patience.

Source: Data Center Dynamics (September 3, 2026)

California lawmakers pass ratepayer protection bills creating separate tariffs for loads above 75 MW

California lawmakers have passed a package of bills intended to stop data center operators from shifting grid and power costs onto residential ratepayers, DCD reported on September 3. The measures now go to Governor Gavin Newsom, who has until the end of September to sign or veto. The package centers on Senate Bill 886, sponsored by state senator Steve Padilla, and Assembly Bill 2383, sponsored by Assemblymember Rick Chavez Zbur. Together they would require the California Public Utilities Commission to establish separate electricity tariffs and updated interconnection rules for large data centers, with AB 2383 dependent on SB 886 becoming law. SB 886, the California Technology Innovation and Ratepayer Protection Act, would require the CPUC to set a tariff covering transmission, distribution and generation costs for new large load customers with peak demand of at least 75 megawatts. AB 2383, the Fair Share in Energy Act, requires separate generation and transmission tariffs for new large-load customers taking service on or after January 1, 2027, with the CPUC given until July 1, 2027 to finalize the structure. The Senate approved SB 886 by 28 to 10 and the Assembly passed it 49 to 7. Padilla also authored SB 887, requiring environmental review under CEQA and giving local communities a formal role in approvals, and lawmakers passed AB 2619 and AB 1577 on water and energy reporting plus AB 2469 on disclosure and infrastructure cost responsibility. The Data Center Coalition opposed the bills, citing privacy and security concerns. Newsom vetoed a narrower water reporting effort last year but told reporters this week that "states all across this country" are leaning in, and that California will too. Similar large-load rate classes are already law in Ohio, North Carolina and Virginia. The practical effect for vendors is a new gate in the schedule. Where a project's economics assumed a general industrial tariff, the pro forma may need rebuilding, and interconnection timing becomes a regulatory question rather than only an engineering one.

Source: Data Center Dynamics (September 3, 2026)

Loudoun County weighs stripping grandfathering from data center applications already in the queue

Virginia's Loudoun County is considering reversing the grandfather clause that protected data center applications filed before the county removed by-right approval, DCD reported on September 3. Under the new rules, data center applications must go through a public hearing and win approval from the board of supervisors rather than being approved administratively. The grandfather clause allowed applications submitted before the change to be evaluated under the earlier regulations, provided they sat more than 500 feet from residential units and were still being diligently pursued during review. On Tuesday, supervisors Juli Briskman and Laura TeKrony proposed reversing that protection. TeKrony argued that Loudoun should not have to "shoulder the burden of all the data centers in the world" and that residents deserve a voice in what is built next to their homes. Supervisor Caleb Kershner argued against, noting that data centers occupy about one percent of county land while contributing roughly half of county revenue. The board did not vote on the reversal itself. It voted instead to have a county attorney assess how many applications are eligible for grandfathering and whether reversing the resolution would be legal, and will wait for that assessment before voting. Nothing is decided, which is exactly why it belongs on your radar now. If you have equipment, long-lead orders or crew commitments tied to a Loudoun project that is relying on grandfathered status, this is the week to confirm with your customer where that application sits and what the fallback schedule looks like.

Source: Data Center Dynamics (September 3, 2026)

Google backs a solar plus long-duration storage project on a reclaimed West Virginia coal mine

Google has partnered with US clean energy developer MN8 Energy and long-duration battery firm Eos Energy on a solar-plus-storage project in Kanawha County, West Virginia, DCD reported on September 3. The Mammoth Solar project combines 86 megawatts of utility-scale solar, 10MW/100MWh of Eos' Z3 zinc-based long-duration storage, and 70MW/280MWh of lithium-ion storage, built on a reclaimed coal mine. Power will be delivered into PJM and support Google's data centers in the region, including a planned West Virginia project. Google will purchase the project's energy, capacity and clean energy attributes. The solar portion is slated for completion in 2028, with lithium-ion storage following in 2029 and long-duration storage in 2030, which would make it the first long-duration storage project in West Virginia. "Because Google was willing to pair next-generation storage with utility-scale solar, we could engineer proven and emerging technologies into one dispatchable resource," said Jon Yoder, president and CEO of MN8 Energy. Lucia Tian, director of advanced energy technologies at Google, described the collaboration as deploying "a smart portfolio of round-the-clock technology solutions." It is also the first joint project under the MN8-Eos master supply agreement signed last October, which targets up to 750MWh of long-duration systems across US renewable projects. Eos is based in Edison, New Jersey and builds zinc-based systems rated for four to 16 hours of duration. Read the staging, not the headline. Three technologies energizing across three different years on one site means three separate commissioning windows, three interconnection milestones and a long tail of integration scope. That is a different opportunity profile than a single-phase solar farm.

Source: Data Center Dynamics (September 3, 2026)

Algoma Steel and Red Jar Energy propose a 400 MW AI data center in Sault Ste. Marie, Ontario

Algoma Steel and Red Jar Energy Partners are planning a 400 megawatt AI data center in the Algoma region of Ontario, DCD reported on September 3. Local outlets published statements from the two companies confirming a proposal in the city of Sault Ste. Marie with a target operational date of October 1, 2029, though no firm commitments have been made. Ron Dizy, managing director of Red Jar, said the companies are seeking an early position in the Independent Electricity System Operator queue so a technical assessment can begin. The project would be subject to local approvals including the recently published Ontario Data Center Playbook, and executives acknowledged that because those guidelines are not yet final, the companies do not know which approval processes will ultimately apply. There is a complication. On Monday, the Sault Star reported that Sault Ste. Marie's city council passed a motion to prepare an interim control bylaw creating a moratorium on data centers until a local framework exists. Algoma Steel, founded in 1901, is an integrated primary steel producer headquartered in the city. "Algoma Steel has shared this city's resources for more than a century," said Laura Devoni, its vice president of human resources and corporate affairs, adding that the company recognizes community questions about electricity, infrastructure and environmental impact. Red Jar developed and built a 100,000 sq ft mining data center facility in 2021, which it sold in 2024. Treat this as an early-stage signal rather than a project. The value in tracking it now is the pattern: heavy industrial sites with existing grid connections and a local employer's political capital are becoming the preferred Canadian entry point, and the IESO queue position is the milestone that separates a press release from a build.

Source: Data Center Dynamics (September 3, 2026)

Hyperscale Data shuts down Bitcoin mining in Dowagiac, Michigan to make room for a neocloud tenant

Hyperscale Data has ended all Bitcoin mining at its Michigan data center as of September 1, clearing the site for a neocloud customer's AI infrastructure, DCD reported on September 3. The unnamed California-based neocloud signed a lease with Hyperscale Data in June 2026. Twenty megawatts are due to be deployed and operational in the fourth quarter of 2026, with the option to scale to 52 megawatts. The initial term runs ten years with two five-year extensions, and at maximum term the contract is valued at roughly 1.2 billion dollars. "The immediate shutdown of the Bitcoin mining operations allows our team to focus the facility's power, infrastructure, and resources in preparing the facility for its usage by our customer," said William Horne, Hyperscale Data's CEO. The facility at 415 East Prairie Ronde Street in Dowagiac spans 617,000 sq ft and was built in 1972 as manufacturing space. Hyperscale Data acquired it in May 2022 and has been repurposing it for HPC and colocation since early 2025, adding plans for a robotics lab in June 2026. The site currently operates at 30 megawatts, and the company aims to expand it to 340 megawatts including 40 megawatts of behind-the-meter natural gas. The company, formerly Ault Alliance, announced a full pivot to AI and data centers in 2024. This is the retrofit trade in its clearest form. A 1972 manufacturing shell with existing power service, going from 30 to 340 megawatts, is a multi-year scope across substation work, gas generation, structural reinforcement and dense liquid-cooled fit-out. Conversions like this rarely get announced as capital programs, so they tend to be underweighted in vendor pipelines relative to the work they actually generate.

Source: Data Center Dynamics (September 3, 2026)

GoPro enters the AI infrastructure market through a merger with Starman Optical

Sports camera maker GoPro is entering the AI data center market via a merger with an optical photonics company, DCD reported on September 3. GoPro disclosed on Tuesday, September 1 that it had entered a definitive merger agreement with Starman Optical and Action Acquisitions. GoPro will remain publicly listed and continue its consumer products, subscription and cloud platform business while adding Starman's US-made optical transceivers, giving it a route into AI infrastructure. GoPro shareholders will receive an aggregate cash payment of 285 million dollars, and the company's 92 million dollars of outstanding debt will be paid in full, leaving a debt-free balance sheet. The deal is subject to shareholder and regulatory approval and is expected to close by the end of calendar 2026. Houlihan Lokey is financial advisor and Fenwick & West is legal counsel to GoPro. "Advanced optics and imaging are essential to AI, national security, and the broader economy, yet much of the critical hardware supporting these technologies continues to be manufactured overseas," said Charles Tebele, CEO of Starman Holding, describing an intent to bring production of these components back to the United States. GoPro founder and CEO Nicholas Woodman said the merger should let the company grow "across consumer, commercial, and defense markets as a leading American imaging and optical solutions company." The transceiver angle is the substantive part. Domestic optical transceiver manufacturing is a genuine supply chain gap in AI cluster buildouts, and onshoring capacity has procurement implications for anyone specifying network hardware under federal or defense-adjacent requirements. Worth noting for context that DCD counts this among a long run of unrelated companies pivoting into the sector, including a wellness retailer, a shoe brand and a whiskey importer, so the diligence bar on new entrants should stay high.

Source: Data Center Dynamics (September 3, 2026)

Velaura AI raises $110 million for ultra-low-power silicon, passing a $1 billion valuation

Silicon Valley-based Velaura AI has raised 110 million dollars in a Series A round that values the company at more than 1 billion dollars, DCD reported on September 3. The round was announced by the company last month and was led by Seligman Ventures, with participation from new investor Capricorn Investment Group and existing investors Mayfield, Maverick Silicon, MARA, Premji Invest, Samsung Catalyst Fund and StepStone Group. Velaura will use the funds to accelerate development and commercialization of its ultra-low-power silicon and software portfolio, including its Titan Core silicon IP and design platform launched in March 2026. The company claims the platform enables up to 2x lower overall chip power for AI accelerators, which it describes as up to 500 watts of savings on a typical 1,000 watt GPU or XPU. "The next era of AI will be defined not only by better models, but also by fundamentally better compute economics," said Rajiv Khemani, co-founder and CEO. Founded in 2022 as Auradine, a Bitcoin mining server company with air, water and immersion cooled products, it rebranded and repositioned to ultra-low-power AI compute in March 2026, having raised 153 million dollars as Auradine in April 2025. Khemani previously founded networking startup Innovium and was COO at Cavium. Treat the 2x claim as a vendor claim until independently benchmarked. The reason to track it anyway is directional. Every watt removed at the chip is a watt not needed in the UPS, the busway, the CDU and the chiller plant, and if efficiency gains of that scale become real in the 2028 to 2030 window, the megawatt-per-rack assumptions inside today's designs will be the thing that ages first.

Source: Data Center Dynamics (September 3, 2026)

Distributed inference: pushing AI workloads out of the mega-campus and into the metro

Equinix announced on September 3 that it is working with Nvidia and Together AI on the Equinix Inference Exchange, described as a distributed AI inference platform. It combines Nvidia's Enterprise Reference Architecture and infrastructure with Together AI's inference platform, delivered through Equinix's global data center footprint and connected via Equinix Fabric to clouds, networks and AI providers. The service is scheduled to be available from the first quarter of 2027.

The architectural argument behind it is worth separating from the marketing. Training wants to be centralized, because it rewards the largest possible contiguous cluster with the cheapest possible power, which is why training capacity keeps landing in Texas, Ohio and Louisiana. Inference wants the opposite. It rewards proximity to the user, to the enterprise's own data, and to whatever jurisdiction the data has to stay in. Equinix operates more than 280 data centers across 77 metros with 230 cloud on-ramps and more than 10,500 interconnected businesses, and the stated use cases follow that logic directly: metro edge inference for latency-sensitive workloads, open model migration for enterprises moving off closed proprietary models, and sovereign workloads that must run in a specific location.

"AI is transforming enterprise technology at extraordinary speed, and the infrastructure decisions enterprises make today will define their competitive position for years to come," said Adaire Fox-Martin, CEO and president of Equinix, describing an architecture that is "neutral by design, open by default." Vipul Ved Prakash, co-founder and CEO of Together AI, framed the aim as removing the trade-off between model choice and performance. Together AI raised 800 million dollars in a Series C in July 2026.

Why it matters for North American builds: if a meaningful share of inference settles into existing metro colocation rather than new greenfield campuses, the work changes character. Instead of one 300 megawatt slab in a rural county, you get dense, high-power retrofits inside occupied, operating facilities across dozens of metros, each with a live SLA, constrained floor loading, existing chilled water and no tolerance for downtime during the upgrade. That is harder work, at smaller unit sizes, with tighter access windows, and it favors contractors and vendors who can execute in occupied space rather than those geared for greenfield volume. It also puts real demand back into markets that lost the training campus race on power price, including the Northeast and coastal metros. Q1 2027 availability means the design conversations are happening in the next two quarters, and the retrofit specifications behind them are being written now.

Source: Data Center Dynamics (September 3, 2026)