The Cloud Goes First: Fourteen Jobs for a Local Node in a Disaster

A 240-watt box with 128 gigabytes of memory keeps answering questions after the WAN, the grid, and the vendor are gone. Here are fourteen jobs it can do — and the limits that decide which ones are real.

Picture the gym at the middle school, eleven hours into an outage. Sixty people on cots. The county’s website has the shelter list, the road closures, the medication-replacement number — and nobody in the room can reach it, because the fiber that serves this valley runs through the same corridor as the fire. The call center that could answer by phone is fielding four times its normal volume. What the room actually gets is a paper flyer with a phone number on it.

This is the part of disaster response that cloud software cannot help with, by construction. Resilience is not a feature you add to a hosted service. It is what remains when the network, the power, and the vendor are all gone.

Which is why a specific piece of hardware has become interesting: a desktop AI machine roughly the size of a hardcover book, drawing about as much power as two laptops, that runs a genuinely capable language model with the network cable unplugged.

Four properties, and why each one matters at hour eleven

We have been sizing projects against the ASUS Ascent GX10, one of several machines built on NVIDIA’s GB10 Grace Blackwell chip. The specifics matter less than the shape of the thing, and four properties do all the work.

• 128 GB of unified memory. Enough to hold a model in the tens of billions of parameters, with room for a long context, entirely on the box. ASUS rates a single node at the 200-billion-parameter class. That is the difference between a toy that needs an API key and something that answers a stranger’s question about their insulin with no internet at all.
• 240 watts peak — measured at roughly 160 watts at the wall under load. Call it 3.8 kWh a day. A node plus a satellite terminal plus a switch lands near 6 kWh a day, which a 10 kWh battery and a properly sized array will carry through a winter storm. No rack server can be islanded on a shelter’s solar. This one can.
• 1.48 kg, 150 x 150 x 51 mm. It fits in a hard case. The compute goes to the incident — the fire camp, the shelter, the Local Assistance Center — instead of the incident trying to send its data somewhere.
• It links to other boxes. Up to three on direct cables, more through a switch. This is the property that turns a box into a mesh, and mutual aid has always been a mesh concept.

The limits, before the list

We put this hardware through an adversarial fact-check and it killed six claims we had been carrying, including two of our own. Publishing the corrected version is more useful than publishing the brochure.

• Memory bandwidth is the ceiling, not the petaflop. The headline “1 petaflop” is sparse FP4; dense is roughly half. What actually limits how fast it answers is 273 GB/s of memory bandwidth. Anyone quoting a petaflop next to a throughput promise is about to be corrected in public.
• CPU and GPU share that 128 GB. NVIDIA’s own reference launch reserves a fraction of it for the GPU. Software that is not written for unified memory will run out well below the number on the box.
• There is one storage slot. Either a 1 TB drive or a 4 TB one — not both, and the cheaper retail unit is the 1 TB. Model weights plus an offline reference corpus plus the operating system do not fit in 1 TB. This single fact decides the bill of materials for half the projects below.
• It computes. It does not transmit. Every one of these projects that reaches an actual resident needs a radio layer — Wi-Fi, LoRa, a ham mesh, satellite. Budget it separately or the field demo fails.
• One box is not resilience. No redundancy, an intake fan, no weather rating. Two nodes is the minimum honest configuration for anything a shelter depends on.

Before the event: detect and prepare

• Ignition watch. Wildfire smoke detection on community-owned cameras, with the inference running locally so detection survives losing the camera site’s uplink.
• Heat outreach. Heat is the leading weather-related killer in the United States on thirty-year averages, and it kills slowly enough to intervene. Cross-reference the forecast against the jurisdiction’s own lists — utility shut-off notices, meal routes, in-home support — and draft the calls. Every one of those lists is sensitive, which is exactly why the work stays in the building.
• Island keeper. Forecast load and solar for the resilience hub itself and set the shed order: what stays powered at hour 12, hour 36, hour 60. The optimizer cannot depend on a service the outage removed.

During: operate with no WAN and no grid

• The blackout desk. A captive portal on the shelter’s own Wi-Fi answering the questions people actually ask — where to go, what road is closed, how to replace a prescription, whether the water is safe — from a corpus held on the box. Voice input, multiple languages, no internet. This is the one that makes the argument visible in thirty seconds.
• Interpreter surge. Telephonic interpreter lines are a single point of failure that fails precisely when demand spikes. Speech translation at shelter intake, running on hardware in the room. Triage and wayfinding only — consequential medical and legal conversations get a qualified human.
• Situation room. Transcribe the radio channels the emergency operations center is authorized to receive, build a searchable timeline, draft the situation report a duty officer currently types at 3 a.m.
• Alert drafting. Turn an incident brief into alert-shaped messages inside the hard character limits, in every language the jurisdiction serves, with a named human approving. A drafting aid that never sends.
• Thin pipe. When the link is 200 bytes and twenty minutes wide, the scarce resource is bytes. Compress field narrative into minimal structured reports, expand them on the far side, hold a queue for the next window.
• Continuity vault. The disaster that actually happens monthly is ransomware or a vendor outage. A clinic or firm keeps a local replica plus a local model, so intake and drafting keep working when the cloud tenancy is unreachable — or encrypted.
• Lifeline ops. A 900-connection water district has one operator and forty years of manuals and valve maps in filing cabinets. At night, with the internet down and the consultant unreachable, the box answers “which valve isolates the school.”
• Shelter class. Schools are the default shelter. The node that runs the district’s tutor on ordinary days becomes the blackout desk when the gym fills. Dual use is the funding argument.

After: recover, document, get paid

• Damage tally. Crews collect windshield and drone imagery; the box classifies it by parcel on site, with no bandwidth to upload and often no connectivity at all. Speed here is money, because the preliminary damage assessment gates the declaration and the declaration gates everything else. Human verification on every parcel that affects a household’s benefit — model output is a triage sort, never a determination.
• Claim clinic. The second wave of a disaster is paperwork. At a Local Assistance Center, help survivors assemble federal applications, insurance claims, loan paperwork and appeals from their own documents — translated, private, never uploaded. Denial and appeal is where recovery money is actually lost, and it is lost hardest by the households with the least paperwork capacity. Survivors are handing over Social Security numbers, deeds, medical records and immigration documents. There is no version of this that runs on someone else’s API and survives a privacy review.
• Mutual aid match. Each town’s node holds its own needs and offers — shelter beds, generators, chainsaw crews, potable water, translators — governed locally, matched across towns, synced when the links come back. The town that is hit is the town whose infrastructure is down, so the matching cannot live in the affected town. It cannot live in a vendor’s cloud either.

Three places this pitch deserves to get pushed back on

Wildfire cameras already have a good incumbent. In California that is ALERTCalifornia, run out of UC San Diego — well over a thousand cameras, free feeds for fire agencies, and in the last reported year it flagged 636 fires before anyone called 911. Against a free statewide system, “we remove the subscription” is not a differentiator. The honest wedge is narrower: detection that survives losing the uplink, and imagery the community owns.

Do not claim federal reimbursement pays for the damage assessment. The management-cost category can generally only be claimed once other public assistance projects are obligated — which requires the declaration the assessment is meant to help obtain. The argument is circular, and a public assistance specialist will catch it in one question.

Wireless emergency alerts are not the multilingual channel. Free-form origination is effectively English and Spanish, and the FCC is explicit that nothing in the chain translates an English message. Multilingual alerting is template-based, with a carrier compliance date in 2028. Point translated output at the systems that can carry it today.

The money is worse than the pitch decks suggest

FEMA’s BRIC program was terminated in 2025, restored by court order, reopened with a billion dollars in March 2026 — and that window closed on 23 July. Future rounds are court-compelled rather than scheduled. The post-disaster mitigation program has its own authorization problems. California’s grid-resilience money is real but attaches to a hub project, not to a compute project.

Which points at the only conclusion in this piece we would defend without qualification. Weight the portfolio toward dual use. A node bought for daily civic and professional work is available in the emergency for free. A node bought only for the emergency is a grant that expires — and when it does, nobody renews the maintenance contract on a box that has been sitting in a closet since the last fire.

That is also why we would start with three of the fourteen and not with a program. The blackout desk as the demo, because a county commissioner understands it immediately. The continuity vault as the revenue, because it sells to people already willing to pay, against a disaster that happens monthly. And mutual aid match as the thesis, because it is the one that turns “buy a box” into “join a mesh” — and the only one on the list a hyperscaler structurally cannot offer.

All three need the same first artifact, and we do not have it yet: a two-node reference deployment with a measured power budget and a benchmarked response time on our own model and our own corpus. Third-party benchmarks for this chip have been public since late 2025. Ours do not exist. Nothing here should be promised to an emergency manager before that changes — and we would rather say so in public than discover it in a shelter.