GridSignal

Talk to enough AI infrastructure operators and a pattern emerges. Whether it's a neocloud, a sovereign AI cloud, an edge AI fleet, or an enterprise running its own cluster, the conversation eventually comes back to the same constraint: Power. The pressure shows up in four ways: ⚡ Can you get enough? New grid capacity and equipment take years to arrive. For most operators, the megawatts you have today are the megawatts you'll have for a while. ⚡ Are you using it well? AI loads jump in milliseconds, while energy and cooling take seconds to minutes to respond. So operators hold power in reserve "just in case." That reserve is paid for and rarely used. ⚡ Is it hurting your reliability? When power or cooling can't keep up, the result is failed jobs, throttled GPUs, tripped circuits, or an edge site that goes dark. ⚡ Is it amplifying your costs? One unplanned peak can set a demand charge that follows you for most of a year. Each operator feels this differently. A neocloud loses sellable capacity. A sovereign cloud has to account for every allocated megawatt. An edge fleet has too many sites to watch one by one. An enterprise finds out about the problem from a breaker or a power bill. GridSignal helps you run more workloads on the power you have by increasing your Ramp Time Lead (RTL), the warning your team gets before a load step lands, and lowering your Stranded Power Ratio (SPR), the share of your power held in reserve but never used. The first step is the same wherever you sit: an X-Ray Report of your own data. gridsgnl.com/x-ray-report #AIInfrastructure #DataCenters #Neocloud #SovereignAI #EdgeAI #PowerManagement

For a decade, the answer to "where should this run?" was simple: the cloud. AI followed the same path. That default is breaking down, and inference is leading the change. As AI moves from experiments into daily production, and from chat answers to agents running inside workflows, enterprises are rethinking where each workload belongs. Three forces drive it: cost that becomes unpredictable at steady volume, data companies won't send elsewhere, and latency that compounds when an agent makes dozens of calls to finish one task. What the data says In Broadcom's Private Cloud Outlook 2026, a survey of 1,800 IT leaders, the share of enterprises running AI inference mainly in public cloud fell from 56% to 41% in a single year, and a Cloudera study found two-thirds have moved some AI workloads back. But this is not an exodus. IDC found only 8–9% of companies plan to leave the cloud entirely. Steady, predictable AI work is moving to infrastructure companies control, while bursty and experimental work stays in the cloud. It's a placement decision, made workload by workload. The four places AI runs now Public cloud, colo and neocloud, on-premises, and the edge: each option trades speed, cost, control and proximity differently, and most enterprises will use all four. The newest option: inference in the building Running AI on the edge means to compute inside or near the buildings where people and data already are. Shah Rahman calls it "NeoEdge": the neocloud wave was built for training, and the next wave will be built for inference. The case is strong. Agents multiply latency, paying the network round trip on every step of a chain. Sensitive data in legal, healthcare and finance wants to stay put. And while new grid connections take years, office and industrial buildings already have power, permits and fiber. The model is still early, but inference is moving toward the people and data it serves. The common thread: every placement is now a power decision Look across the four options and one constraint shows up in all of them. Cloud providers are limited by how fast utilities can connect new campuses. On-premises teams find a GPU rack draws more power than their facility was built to deliver. Colo and neocloud capacity is only as available as the provider's power contracts. And the edge depends on how much a building can really spare, which is far less on a hot weekday afternoon than on paper. AI is coming home. Wherever it lands, it brings its power demands with it. Sources Broadcom, Private Cloud Outlook 2026 (summary via LLM Configurator) Cloudera survey of 1,500 enterprise technology leaders, Virtualization Review, Aug 2026 IDC repatriation research, 2024 (summary via Abhishek Kushwaha) Shah Rahman, "NeoCloud Was for Training. NeoEdge Is for Inference," Jul 2026

Claude Code and Replit didn't teach me how to build GridSignal's simulation engine. Learning to write functional and design specs, write and review modular code, and run unit, black box, and integration tests did that. AI hasn't changed software development, but AI has certainly accelerated software development.

I’m excited to announce that I’ll be attending the inaugural #NovaFutureSummit26 this September 28-30 in Napa, California. The summit presents a unique opportunity to gather with other leaders in tech, telecoms and business for substantive, closed-door discussions on the future of AI and global digital infrastructure. If you'll be in Napa, I'd love to connect. #novafuturesummit26

GridSignal is opening our Founding Design Partner Program! GPU clusters can ramp from idle to full power in seconds. Most facilities only find out when the meter does. LivePulse reads the job scheduler to forecast those power ramps before they happen, giving operators time to prepare generation, storage and cooling. It's read-only and advisory by design. We never touch your workloads or equipment. We're inviting a small number of founding partners into our pilot. If you run dense GPU clusters on Slurm, Kubernetes or Ray and have access to your facility's power data, we'd love to talk. Founding partners receive a complimentary X-Ray Report on their historical data, direct access to our engineering team, and a real voice in what we build. gridsgnl.com/newsroom/beta-p…

Curtailment Mitigation for PCLRs: Modeling the Effects of Front-of-Meter Netting published byAurora Energy Research @AuroraER_Oxford auroraer.com/resources/auror…

Your job scheduler already knows a power spike is coming — 30 to 60 seconds before your meters do. That's not a marketing number. It's a physical one. When a training job or a large inference batch gets committed to the queue, that's a decision that's already been made. The power draw that follows is a consequence, not a surprise — it just hasn't happened yet. Meanwhile, the systems that actually respond to that draw run on their own clocks. Generators and switchgear can ramp in seconds once triggered. Cooling systems notice the heat rise on a slower, separate clock. Today, almost everything downstream waits for the meter to see the spike before either clock starts moving — reacting to something that was already knowable. That gap between "the scheduler committed the work" and "the meter saw the draw" is the whole opportunity. It's advance warning that already exists in data centers' own systems, going unused. That's the gap GridSignal reads. gridsgnl.com #DataCenters #AIInfrastructure #PowerManagement

Press Release Stats - GridSignal Launches Free Diagnostic to Reveal Hidden Power Capacity in AI Data Centers prlog.org/13170836-gridsigna…

For 50 years, power demand held roughly flat. Then AI broke that. Most data center teams find out about a power spike the same way they always have: after it happens. But the warning was often already sitting in their own job scheduler, unread. We built the GridSignal X-Ray Report to answer one question with evidence instead of guesswork: how much advance warning did your power infrastructure already have, and what did it cost you to miss it? It's a free, one-time diagnostic. No integration, no system access — you send us 90 days of your own scheduler and power data, and we compute six figures: lead-time opportunity, capacity utilization, peak events, demand-charge ratchet risk, recoverable stranded capacity, and the full method behind every number, written out for your engineers to check. Evidence before assumption. That's the whole idea. If you run AI infrastructure and want to see what your data is already telling you, request your complimentary X-Ray Report here. gridsgnl.com/x-ray-report