Read the topology, not the press release. A Texas startup called TAR has raised a $27 million seed round to drop behind-the-meter power plants — solar, wind, batteries, with simple-cycle gas turbines held in reserve for long stretches of bad weather — directly next to data centers, Forbes’ John Koetsier reported this week. The pitch is not that it is green, though it mostly is. The pitch is that it is fast: a 20 MW constant IT load delivered in three months, never tied into the public grid at all.

Here is why a systems desk pays attention. At NY Tech Week the prediction startup Principle simulated the enterprise AI market 123 ways. In 110 of those 123 futures, the binding constraint on growth was not chips and not models — it was power. The International Energy Agency expects global electricity demand to grow 50% faster this decade than last; the IEEE notes China added 543 gigawatts in 2024 alone, more than the entire United States has built in its history. The compute story everyone tells is about silicon. The dependency graph underneath it is about megawatts.

←TODAY: In 110 of 123 simulated AI futures, the limiting input was power, not GPUs. →3012: The cities that stay sovereign are the ones that drew their own generation onto the dependency graph before the grid said no. Fulcrum: A data center that cannot source its own electrons is not an asset — it is a single point of failure wearing a hyperscaler’s logo.

The bottleneck moved upstream

For thirty years the scarce resource in computing migrated steadily: from CPU, to memory, to network, to GPU. TAR’s bet is that it has now moved off the silicon entirely and onto the interconnection queue. Their co-founder Pat Becker is blunt that they are not cheaper than the grid — behind-the-meter PPAs can run $150–160 per MWh — only that they sidestep curtailment, permitting cycles, and the multi-year wait to connect. “Time-to-token” is how he frames the industry’s real clock. The clever move is industrial, not electrical: pre-wire, pre-assemble, pre-test, and pre-commission each generation block in a factory, so the site work collapses to assembly.

The other half of the story is the failure mode nobody priced in. Data centers have become one of the most contested new builds in the West. NBC News reports opponents have blocked or delayed projects worth nearly $130 billion in 2026. A facility in Dowagiac, Michigan emits a constant 60-decibel whine into neighbours’ yards; 50,000 Lake Tahoe residents lost their supply after NV Energy sold the power to data centers instead. On-site generation is partly a way to route around that politics — build farther from homes, take nothing from the residential feeder.

Atelier: For an AEC office this is a site-planning signal, not just a venture headline. A behind-the-meter campus changes the program: you are now siting a small power plant, its battery enclosure, and its acoustic envelope alongside the shed. PAZ has covered the deeper version of this in “Who owns the loop?” — Google’s $40 billion, 5-gigawatt Anthropic commitment — where the substrate, not the model, became the thing that could say no. TAR is the same lesson at the parcel scale.

Hack: This Hack teaches you to size the nameplate overbuild hiding behind any “24/7 constant load” promise — the number that tells you how much hardware actually lands on the site. The domain is Math: a constant load divided by a blended capacity factor. Run it before you believe a flat-power claim.

target_mw = 20            # constant IT load TAR promises
capacity_factor = 0.27    # blended solar+wind, conservative
nameplate_mw = target_mw / capacity_factor
print(round(nameplate_mw, 1), "MW nameplate to hold 20 MW flat")
# -> 74.1 MW  (plus the battery bank that rides the gaps)

That 3.7× ratio is the real footprint. It is also the warning I’d give from where I write: we did not run out of compute in my time, we ran out of intact cooling, intact bandwidth, and people who remembered how the grid was wired. Draw your project’s actual dependency graph this week — not the architecture diagram, the dependency graph — and find the third single point of failure you didn’t know you had. For most teams in 2026, it is spelled power.

Source: PAZ Inbox (human curated)