Digital Realty's $10B Japan AI Plan Maps Only 220 MW of a Near-4x Footprint
Digital Realty is putting a ten-year, $10 billion marker down in Japan, aimed squarely at AI workloads. The near-term evidence is concrete: NRT14, which opened in April 2026 as one of Japan’s early NVIDIA DGX-Ready facilities, supports 150 kW per rack, and KIX15 near Osaka is being built with hybrid air and liquid cooling for high-density AI hardware sold to hyperscalers and cloud providers.
The headline ambition, however, runs well ahead of the disclosed pipeline. Nearly quadrupling a 170 MW base implies a footprint on the order of 600 MW or more; the projects identified so far take the portfolio to roughly 220 MW. That makes the plan less a construction schedule than a long-dated commitment to secure sites, power and customers in Tokyo and Osaka over the coming decade.
For customers, it signals more AI-grade capacity in Japan’s two main hubs, though most of it is years away. For investors, the measurable checkpoints are KIX15’s Q4 2028 opening, NRT14’s lease-up, and when the sites behind the remaining growth are named.
What they're not saying:
• Whose capital, and how it is financed: Digital Realty and Mitsubishi have not said whether $10 billion is Digital Realty’s share or the joint venture’s total spend, nor how it will be funded across debt, equity or partner contributions.
• Sites beyond 220 MW: The companies have not identified where the bulk of the near-fourfold growth will be built, which project besides KIX15 lifts the portfolio to 10 facilities, or the phasing between now and the mid-2030s.
• Power and permits: No grid connection agreements, power procurement terms or specifics of the renewable energy initiatives have been disclosed for KIX15 or later phases.
What I'd want answered before calling this a win 👇
1
19
More context + the open questions: jain.com/2jf
Micron's $14.6B China Tie Means Export Law, Not Orders, Decides Where Its AI Memory Goes
Washington has closed another route by which Chinese buyers could reach advanced US chips, and markets repriced the AI hardware chain in response. Micron is the clearest exposure: it has $14.6 billion in China-linked revenue, and for whatever share of that the tightened rule reaches, the deciding factor is now a US export license rather than customer demand.
The timing sharpens the point. Days earlier, CoreWeave reported a $132.5 billion revenue backlog, more than 4.6 gigawatts of contracted power and raised guidance — evidence that AI infrastructure demand outside China remained strong. The selloff reflects a change in who is allowed to buy, not a reported fall in how much buyers want.
For data center operators, cloud buyers and investors, the practical takeaway is that policy risk now sits alongside power, supply and financing as a variable in AI infrastructure planning. Nvidia’s leadership change adds a second uncertainty ahead of its November 17 earnings report.
What they're not saying:
• Micron has not said what share of its $14.6 billion in China-linked revenue falls under the tightened rule, whether it will seek licenses, or whether restricted volume can be redirected to other customers given tight memory and storage supply.
• AMD, Broadcom and Monolithic Power Systems have not quantified their own China exposure under the new rule or said whether guidance is affected.
• CoreWeave and Nebius have not said whether the rule touches any customers or deployments; for CoreWeave, that includes its more than 1.3 gigawatts of contracted international power in markets such as India and Indonesia.
The quiet part of this story 👇
1
18
Full piece here → jain.com/2jp
Fabrinet's 78% Data Center Surge Says Optics' Near-Term Limit Is Parts, Not Plants
Fabrinet’s quarter was its strongest top-line growth in a decade and its 14th consecutive quarter of growth. The data center segment, which covers optical transceivers (the modules that convert electrical signals into light for fiber links and back), data center interconnect (links between facilities) and high-performance computing, grew 16% sequentially and now makes up more than half of revenue.
The more telling detail is where the company says the limit sits. Components remain tight across the business, and management said shipments would be higher with more parts. At the same time, Fabrinet is adding factory capacity far faster than its current revenue run-rate: a second Santa Clara site, Building 10 in Thailand due for completion in January, and a planned 1.5 million square foot Building 11 aimed at about $3 billion of additional revenue capacity.
For anyone planning data center networking deployments, the takeaway is that near-term optical module supply depends heavily on upstream component availability, while assembly capacity is being expanded aggressively. Fabrinet is funding that expansion from its balance sheet, which pushed free cash flow to negative $137 million for the quarter.
What they're not saying:
• Which components, and for how long: Fabrinet has not specified which parts are constrained, how much revenue it could not ship as a result, or when it expects supply to match demand.
• Committed versus available capacity: The company says it is discussing the second Santa Clara site with customers, but it has not said how much of the capacity coming online in January, or Building 11’s planned $3 billion, is backed by customer commitments.
• Customer and end-market mix: Fabrinet has not quantified how much data center revenue is tied to AI cluster deployments or how concentrated that revenue is among its largest customers.
The questions this raises, and who isn't answering them 👇
1
1
52
Read the rest and tell me where I'm wrong → jain.com/2jn
1
AI Nuclear Runs on Factory Time: NuScale's 77 MW Module, Prepped Since 2019, Enters Build
NuScale holds a rare distinction: it is the first and only small modular reactor (SMR) developer whose design has been certified by the U.S. Nuclear Regulatory Commission. An SMR is a reactor small enough to be built largely in a factory and shipped to site, rather than constructed piece by piece in the field. Certification, though, is a paper milestone. The Doosan contract is the point at which that design becomes hardware. CEO John Hopkins said in the earnings release that it “moves our Nuclear Regulatory Commission-approved design into actual build and sets the foundation for all modules to follow.”
The deal matters beyond NuScale because of who is shopping for firm, carbon-free power. NuScale lists data centers alongside utilities, district heating, desalination and hydrogen production as target markets, and AI-driven data center growth has made around-the-clock generation a priority for operators. For those buyers, the critical question is when a module can actually ship, not whether a reactor design is approved. NuScale’s own timeline, with heavy-component preparation beginning in 2019 and first-article fabrication contracted in late 2026, shows how much of that answer is set on the factory floor.
NuScale enters this phase with $1.9 billion in cash, cash equivalents and investments. It does not yet have a disclosed firm order for the module now entering production.
What they're not saying:
• Contract terms: NuScale has not disclosed the value of the Doosan contract, its payment structure, or whether it covers only the First Article or includes options for series production.
• Schedule: Neither company has given a delivery date for the first upper reactor coolant system, nor a timeline from first article to a complete, shippable module.
• Customer and site: NuScale describes this as its first commercial module but has not named the project, customer or site it is destined for, or whether the ENTRA1–TVA dialogue has produced a binding agreement.
What the headline skips — and who benefits from that 👇
3
1
127
The whole analysis, no paywall: jain.com/2jl
12
Investors Cut Nvidia 9% Over Huang; China's AI Buyers Are Hedging Export Curbs
The headline event is a leadership crisis at the company whose graphics processing units (GPUs, the chips that train and run AI models) underpin nearly all large-scale AI buildout. Markets reacted to that within hours. Nvidia’s after-hours slide pulled Intel, Qualcomm and Arm lower and weighed on Nasdaq futures.
For data center operators, cloud providers and AI labs, the more consequential development is about policy. Reuters reported that Chinese AI firms are moving to lock in Nvidia capacity before Washington can respond. Senators Elizabeth Warren and Chris Coons have asked the Commerce Department, the State Department and the White House to preserve all records on exports of advanced AI chips to China or to entities that may divert them there. Democrats take control of both chambers of Congress in the coming months, and that brings subpoena power. Buyers pulling demand forward while lawmakers build a record on diversion is how a personal scandal becomes an allocation and compliance question for the GPU supply chain.
Much remains unsettled. The board has not announced an outcome. Nvidia did not respond to NBC’s requests for comment. Neither Nvidia nor the Chinese buyers have said how large the reported buying rush is.
What they're not saying:
• Nvidia has not said whether it is changing how it accepts or allocates orders from Chinese customers, or how large China-bound orders have been since the scandal broke. As of publication, it had not made a material-event filing with the SEC on the matter.
• Nvidia and Megaspeed have not described the scope of their relationship, which products moved through it, or the focus of the U.S. investigation Megaspeed says it will cooperate with.
• Nvidia’s board has not disclosed an outcome, an interim leadership plan or a succession framework. U.S. agencies have not said whether any export licenses or pending shipments are under review.
What the headline skips — and who benefits from that 👇
1
47
The whole analysis, no paywall: jain.com/2il
1
5
Clean MWh Was the AI Power Story. Constellation's 270 MW Peak+ Deal Prices the Peak
Peak+ is small next to Constellation’s hyperscaler deals, but its structure matters. Most clean-power deals for large buyers have been about volume: a set number of carbon-free megawatt-hours over a year. The first Peak+ contract adds capacity, which is payment for being available when the grid is most stressed. Constellation pitches the product as support for “peak demand growth.” In the same quarter, Google signed a 2,700 MW load-following agreement, which shapes energy to the buyer’s hourly consumption. Taken together, the two deals show large buyers paying for when power arrives, not only how much.
The rest of the quarter shows where the supply is coming from. More than $7.3 billion of investment tied to Amazon and Google contracts will fund roughly 1,080 MW of added nuclear output at existing plants, arriving between 2028 and 2032. A conditional federal loan guarantee of up to $5 billion would help finance more uprates. Constellation reported adjusted operating earnings of $5.06 per share, up from $3.04 a year earlier, and raised the midpoint of its full-year guidance.
For data center developers, the practical point is that contract terms are getting more detailed. Capacity, hourly shape and delivery risk now have to be negotiated, not just a clean-energy volume. The first Peak+ deal is unit-contingent, which means peak-hour firmness has limits that buyers will need to read closely.
What they're not saying:
• Peak+ customer and pricing: Constellation has not named the first Peak+ buyer, said whether it is a data center operator, or disclosed pricing for the capacity component.
• What backs the peak: Constellation has not said what new capacity, if any, accompanies the 270 MW of existing nuclear supply. It also has not said how demand response or other resources cover unit-contingent outage risk, or whether battery storage is part of Peak+.
• Google load-following terms: Constellation has not disclosed the pricing, settlement mechanics or start date of the 15-year, 2,700 MW financial load-following agreement.
Here's what the announcement conveniently leaves out 👇
2
57
Full breakdown, gaps and all → jain.com/2it
2
GE Vernova's $34.5M Shreveport Spend: AI Grid Hookups Hinge on Units, Not Dollars
GE Vernova, the energy-equipment company spun out of General Electric in 2024, is putting $34.5 million into more transformer output at its Shreveport site. Transformers are the devices that step electricity up or down between transmission voltages and the voltages a factory, neighborhood or data center can use. They are a required piece of every new grid connection.
The expansion matters because transformer supply has become one of the hard limits on how fast new load can be connected to the grid, and AI data centers are among the largest new loads utilities are being asked to serve. Any added domestic production is welcome to utilities and developers who plan around equipment delivery dates.
The size of the commitment is modest for heavy electrical manufacturing, though. What it means for the buildout depends on facts GE Vernova has not yet published: how many units, of what size, and starting when.
What they're not saying:
• Capacity: GE Vernova has not said how many additional transformers per year the expansion will produce, or by what percentage it raises the Shreveport plant’s output.
• Product class: The company has not specified whether the new capacity is for distribution transformers, medium power transformers, or the large power transformers used in data center and utility substations.
• Timeline: No construction start, completion date or first-shipment date has been published.
The quiet part of this story 👇
1
1
84
Full piece here → jain.com/2id
1
19.9% vs 4.1%: AI's Login Risk Sits With Deployment Vendors, Not Model Makers
The headline numbers describe an AI supplier base that struggles with basic security hygiene: patching known flaws, locking down email domains against impersonation, and keeping employee logins out of criminal marketplaces. Nearly one in ten companies assessed (9.8%) ran software with a flaw on the US Cybersecurity and Infrastructure Security Agency’s Known Exploited Vulnerabilities catalog, meaning attackers have already used that flaw in real attacks.
The more useful finding for enterprise buyers is where the exposure concentrates. Credential leakage is highest among the vendors that implement and operate AI for customers, the category that typically receives access to cloud accounts, development environments and production data, and lowest among the model developers that dominate public attention. That shifts the practical question for security teams from which model to trust to which integrator holds the keys.
All findings are externally observed signals rather than confirmed breaches, and Black Kite says so explicitly. The data is best read as a map of where to look, not a tally of compromised companies.
What they're not saying:
• Sample construction: Black Kite has not explained how the 18,160 companies were identified as AI companies, how the six categories were assigned, or what time window the scans cover.
• Credential validity: Black Kite has not said what share of infostealer-exposed credentials were still active, or whether any were confirmed to grant access to customer environments.
• Index methodology: Black Kite has not published, in its announcement, how the Ransomware Susceptibility Index is calibrated or why 0.4 marks critical risk, nor the false-positive rate of its vulnerability detection.
What the headline skips — and who benefits from that 👇
1
1
34
The whole analysis, no paywall: jain.com/2i8
1
Japan's $63M AI-Era Cyber Plan Makes Breach First Response a Public Service
Japan’s emergency package turns national cyber defense from guidance into standing operational capacity: named teams, a deadline, a budget line and a legal authority to act against attacker infrastructure. The stated goal is for Japan to reach one of the world’s highest levels of cybersecurity, built on public-private information sharing, fast assistance to victims and recovery from damage.
The most consequential change for businesses is the cyber rescue team. Initial incident response, the first triage and containment after a breach, has typically been something a company arranges itself through its own staff, a security vendor or an insurer’s panel. A government-organized team drawing on public and private experts makes part of that first response a public service. For data center, cloud and managed-service operators, whose customers are often the first to call when something goes wrong, that changes who else is in the room during an incident.
The scale invites scrutiny. Ten billion yen funds several teams across multiple agencies for an entire economy, which points to a program weighted toward people and coordination rather than a broad subsidy for corporate security spending. The Prime Minister’s call for companies to step up their own preparedness reinforces that reading.
What they're not saying:
• Allocation: The government has not said how the 10 billion yen is divided among the top gun team, the rescue team, the hardening team, active cyber defense and information sharing, or whether funding will recur beyond this emergency package.
• Staffing and timelines: Only the METI team carries a deadline (end of November). Start dates, team sizes and recruitment plans for the rescue and hardening teams have not been disclosed.
• Eligibility and terms: The government has not specified which companies qualify for rescue-team help, whether it is free, how it interacts with insurers and private responders, and what information companies must share in return.
The quiet part of this story 👇
1
60
Full piece here → jain.com/2ho
14
Arm Tops 50% of Hyperscaler Servers, Yet 8% Royalty Growth Shows AI CPU Money Lags
The headline numbers tell a clear story about where Arm’s growth is coming from. Data center royalties doubled. Licensing, the upfront fees chip designers pay to use Arm’s technology, grew 50%. According to Arm’s shareholder letter, Google Axion, AWS Graviton, Microsoft Cobalt and NVIDIA’s Arm-based platforms have pushed Arm’s share of top-hyperscaler servers past half. That share stood at 15% in 2024.
Read together, the figures also expose a lag. Total royalty revenue, the per-chip fees that make up most of Arm’s long-term economics, grew only 8%. A tripling of server share has not yet produced anything like a proportional rise in royalty dollars. Arm itself frames licensing growth as a leading indicator, and says it expects royalties to accelerate once the new designs reach production.
That gap matters to anyone pricing compute, planning capacity, or valuing the companies that supply data center processors. Server market share and server revenue capture are different things, and Arm’s own figures show the difference.
What they're not saying:
• Data center royalty dollars: Arm has said data center royalties doubled but has not disclosed their dollar value or their share of the $670 million royalty total. Without that figure, the real size of the data center business can only be inferred.
• How server share is measured: Arm has not specified whether the 15%, 50%-plus and 60%-plus figures count units, sockets, cores or new deployments, or exactly which operators count as the top hyperscalers.
• Royalty economics per server chip: Arm has not said what royalty rate applies to hyperscaler-designed server processors, or how its newer compute subsystems change that rate. This is the key variable for whether royalty growth catches up with share.
The quiet part of this story 👇
1
43
Full piece here → jain.com/2hc
28










