The Data Center in the Basement: Why Micro Home Compute Could Break AI’s Capacity Bottleneck
A cost model, a capacity model and runnable code for engineers and investors deciding whether residential compute nodes are a real answer to AI’s power shortage or a clever press release.
The binding constraint on AI compute in 2026 is not silicon. It is a powered building. In November 2025 Satya Nadella said Microsoft may have “a bunch of chips sitting in inventory that I can’t plug in,” because the company lacked energized shells to put them in. Transmission queues, substation transformers and utility studies now set the pace of AI capacity, and they move in years while GPU generations move in months.
That gap is why an electrical panel company, a GPU vendor and a homebuilder announced, in the spring of 2026, a plan to bolt 16-GPU, liquid-cooled inference nodes onto the outside walls of new American houses. SPAN’s XFRA program, built with NVIDIA and PulteGroup, targets a 100-home pilot this year and more than 1 GW across roughly 80,000 homes in 2027. In the UK, Heata and Thermify are already running servers that heat water and homes. The idea has moved from hobbyist novelty to a financed deployment plan.
My argument in this piece is narrower than the headlines and, I think, more durable. Micro home data centers do not beat a hyperscale campus on unit cost. On my numbers they cost about 7% more per GPU-hour at equal utilization. What they beat it on is time to power, measured in months instead of years, and when compute is scarce, time is the expensive input. Below you will find the framing model, a cost model you can run, the bandwidth math that decides which workloads fit, a transformer-headroom simulation that shows where the idea breaks, and a sourced list of every number I used.
The bottleneck moved from chips to wires
Demand is outrunning the grid’s ability to connect new load. Lawrence Berkeley National Laboratory put US data center consumption at about 176 TWh in 2023, 4.4% of national electricity, and projected 325 to 580 TWh by 2028, or 6.7% to 12%. Divide by 8,760 hours and that is a move from about 20 GW of average load to somewhere between 37 and 66 GW, which means adding roughly 3.4 to 9.2 GW of new average load every year. The IEA projects global data center demand more than doubling to about 945 TWh by 2030, with demand from AI-optimized facilities more than quadrupling.
The supply side cannot absorb that pace through the normal channel. LBNL’s Queued Up report counted 2,061 GW of generation and storage still seeking interconnection at the end of 2025, with a median of more than five years from request to commercial operation. Of the capacity that requested interconnection between 2000 and 2020, only 13% had been built by the end of 2025. In PJM, the largest US grid operator, projects entering service in 2025 averaged more than seven years to reach operation, according to PJM data reported by Data Center Knowledge (PJM disputes how some analysts count that timeline).
The hardware between the substation and the server is short as well. Wood Mackenzie data show average transformer lead times rising from about 50 weeks in 2021 to about 120 weeks in 2024, and substation transformers now exceeding 160 weeks. PwC analysts put large power transformers at up to four years, with prices up roughly 80% over five years.
The market is pricing the scarcity. Average H100 rental rates rose from about $1.70 to $2.35 per GPU-hour between October 2025 and March 2026, roughly 40%, per SemiAnalysis data reported by Crypto Briefing, after two years of falling prices. Retail electricity followed: the EIA expects the average US residential price to reach 18.2 cents per kWh in 2026, against 14.0 cents for commercial and 9.05 cents for industrial customers, and it attributes part of the rise to commercial demand that includes data centers.
A new 100 MW campus waits years for a grid connection that is sized, studied and built as one large point of load. The distribution network that already reaches every house is sized for peaks that occur a few hours a year. SPAN’s CEO Arch Rao argues that the distribution network runs at only 40% to 45% utilization on average. I could not find an independent study that confirms that number, so treat it as a vendor hypothesis, but the direction is consistent with how residential feeders are designed: for the evening peak, not the average.
The framing model: cost per sold GPU-hour, and the value of being early
Two equations organize everything that follows. The first is the all-in cost of one GPU-hour that a customer actually pays for. It is a proposed model, deliberately simple, and it works for any site type.
c = [K_IT·CRF(r, n_IT) + k_fac·P·CRF(r, n_fac) + O + N + S − H] / (8760·u) + P·PUE·p_e
Here K_IT is the capital cost of a GPU plus its share of the server, k_fac is facility or enclosure cost per kW of IT load, and P is IT power per GPU in kW. CRF(r, n) = r(1+r)^n / ((1+r)^n − 1) is the capital recovery factor, which turns an upfront cost into an equal annual charge at discount rate r over n years. O is operations and maintenance per GPU-year, N is network cost, S is any payment or subsidy to the site host, and H is a credit for useful heat. The variable u is utilization, the fraction of the year’s 8,760 hours that are sold. PUE is power usage effectiveness and p_e is the electricity price per kWh. Energy sits outside the fraction because a GPU only burns full power while it is working.
The structure tells you what matters before any numbers go in. Every fixed cost is divided by u, so utilization is a multiplier on everything except energy. A site can win on electricity, cooling and construction and still lose if it cannot keep its GPUs busy.
The second equation is the value of speed. Suppose a micro fleet can energize capacity at time T_m and a campus at T_h, with T_m < T_h. During that window only the fast builder can sell compute. For G GPUs earning market price p against cost c, the present value of the head start is:
V = ∫ from T_m to T_h of (p − c)·8760·u·G·e^(−r·t) dt
This term is usually ignored in data center cost comparisons, which assume both options exist at the same moment, an assumption that fails when compute is scarce. The rest of the piece puts numbers into both equations.
Who is already doing it
Three distinct models are in the field, and they differ more than the shared “data center in your home” headline suggests.
SPAN XFRA, built with NVIDIA and PulteGroup, puts a liquid-cooled cabinet on the exterior wall of a new home, along with a smart panel and a home battery. Each node holds 16 RTX PRO 6000 Blackwell Server Edition GPUs and 4 AMD EPYC CPUs. Installation is free, and according to Latitude Media the host pays about $150 a month while SPAN covers the home’s electricity and internet. The pilot is 100 homes in 2026, about 1.25 MW and 1,600 GPUs, and SPAN’s announcement targets more than 1 GW in 2027.
Heata, in the UK, straps a server unit to the household hot water cylinder and runs general cloud workloads on it. Heata pays the unit’s electricity, and each unit supplies up to 4 kWh of hot water a day. Fortune reports about 100 homes and about 1 GWh of energy saved, and British Gas ran a 10-unit trial.
Thermify’s HeatHub replaces the boiler with an oil-immersed cluster of Raspberry Pi compute modules running containerized jobs, as part of UK Power Networks’ SHIELD trial for low-income tenants. One Essex household’s bill fell from £375 to as little as £40 a month, TechSpot reported.
Prime Intellect’s INTELLECT-2 used contributors’ own GPUs, with no permission needed to join, to generate reinforcement learning rollouts for a 32B model. Rollouts ran on untrusted nodes; the training updates stayed on trusted ones.
SPAN’s design is the one that matters for AI capacity. Each node carries 1.5 TB of GPU memory (16 cards at 96 GB), which is enough to hold a 671-billion-parameter mixture-of-experts model at 8-bit precision with room left for KV cache. The pilot figures imply about 12.5 kW per home, and SPAN caps each node at 80 amps, which is 19.2 kW at 240 volts, on homes kept at a standard 200-amp service. SPAN says it owns the compute and batteries and sells capacity to hyperscalers, neoclouds and AI companies.
The UK systems solve a different problem. They are heat products that happen to compute, sized in hundreds of watts per home, and the economics rest on displacing a gas or electric water heater. That distinction comes back in the section on where the model breaks, because heat reuse does not scale to a 12.5 kW GPU node.
The last example is a reminder that the software side is further along than the hardware. INTELLECT-2 showed that a 32B reasoning model can be trained with reinforcement learning where the rollout generation, which is just inference, runs on a heterogeneous swarm of consumer GPUs that join and leave at will, which is the pattern a fleet of home nodes would serve.
The cost model: a near tie on unit cost
On my assumptions a home node costs $1.22 per sold GPU-hour against $0.93 for a hyperscale campus, but most of that gap comes from assuming the home fleet runs at 60% utilization against 75%. At equal utilization the gap shrinks to about 7%. I compare the same GPU in both places, an RTX PRO 6000 Blackwell Server Edition, so the comparison isolates the site. Every input is listed below with its basis; the ones marked as assumptions are mine, and you should replace them with your own.
Input Campus Home node GPU + server share $15,000 $15,000 Facility, per MW of IT $12M $3M Facility life 20 yr 10 yr IT power per GPU 0.72 kW 0.72 kW PUE 1.25 1.08 Electricity 9.05¢/kWh 18.2¢/kWh O&M per GPU-year $600 $1,200 Network per GPU-year included $112.50 Host subsidy per GPU-year $0 $56.25 Utilization 75% 60% Discount rate, GPU life 10%, 5 yr 10%, 5 yr
Where the inputs come from: the GPU price of about $10,000 is from TechRadar, and the server share is my assumption. The campus facility cost sits near JLL’s shell-and-core average of about $11.3M per MW, via CORE Advisors; the home figure of about $3M per MW is SPAN’s own claim. Electricity prices are the EIA’s 2026 industrial and residential forecasts. The 0.72 kW per GPU comes from 12.5 kW per 16-GPU node, including CPUs and pumps. PUE, facility life, O&M, network ($150 a month of fiber per node), host subsidy ($75 a month net after the host’s $150 fee), utilization and discount rate are my assumptions; the O&M figure assumes truck rolls and spares double the field cost.
The result, per sold GPU-hour, with components rounded:
Per sold GPU-hour Campus Home node GPU and server capital $0.602 $0.753 Facility or enclosure $0.154 $0.067 Electricity $0.081 $0.142 Operations and maintenance $0.091 $0.228 Network and host $0.000 $0.032 Total $0.930 $1.222
The GPU is about 62% of the home node’s cost and 65% of the campus’s, so siting only moves the remaining third. The home node wins on the building by $0.09 per hour and loses on electricity by $0.06 and on field operations by $0.14. Much of the total gap comes from the utilization assumption. Run the home node at 75% and it costs $1.006, 8% above the campus. To match the campus outright it needs 82% utilization, or about 75% if it gets a commercial tariff of 14 cents and holds O&M to $900.
The comparison that matters to an operator, though, is against the price of compute, not against a campus that does not exist yet. RTX PRO 6000 capacity rented for $1.64 to $5.50 per GPU-hour in October 2026 across nine providers tracked by Thunder Compute, with Vast.ai at the bottom, Nebius at $1.80 and CoreWeave at about $2.50. At $1.64 the home node breaks even at 43% utilization, at $1.80 at 39%, and at $2.50 at 27%. That is a comfortable margin of safety if demand holds.
[embed: node/b96510df-ca53]
The two cost curves run almost parallel, about 8 cents apart. Whether either one clears the market depends far more on utilization than on location.
Now the second equation. At a price of $1.80, the home node earns $0.58 of margin per sold hour, or about $3,040 per GPU-year at 60% utilization. One MW of IT load holds about 1,389 of these GPUs, so that is about $4.2M per MW-year. If the home fleet is energized at six months and the campus at three years, which is the low end of SPAN’s three-to-five-year claim and well short of PJM’s recent average, the head start is worth about $8.9M per MW in present value. That is close to the $12M per MW the campus spends on its building.
from dataclasses import dataclass, replace
import math
def crf(r, n):
"""Capital recovery factor: upfront cost -> equal annual charge."""
return r * (1 + r) ** n / ((1 + r) ** n - 1)
@dataclass(frozen=True)
class Site:
name: str
k_it: float # $ per GPU, GPU plus its share of the server
k_fac_per_mw: float # $ per MW of IT load, facility or enclosure
fac_life: int # years
p_it_kw: float # IT load per GPU, kW
pue: float
price_kwh: float # $ per kWh
om: float # $ per GPU-year
net: float # $ per GPU-year
host: float # $ per GPU-year paid to or for the host
util: float # share of hours sold
heat: float = 0.0 # $ per GPU-year heat credit
r: float = 0.10
it_life: int = 5
def cost_per_hour(self):
hours = 8760 * self.util
fixed = (self.k_it * crf(self.r, self.it_life)
+ self.k_fac_per_mw * self.p_it_kw / 1000 * crf(self.r, self.fac_life)
+ self.om + self.net + self.host - self.heat)
return fixed / hours + self.p_it_kw * self.pue * self.price_kwh
def breakeven_util(site, price):
lo, hi = 0.01, 1.0
for _ in range(60):
mid = (lo + hi) / 2
lo, hi = (mid, hi) if replace(site, util=mid).cost_per_hour() > price else (lo, mid)
return hi
def speed_value_per_mw(site, price, t_fast, t_slow):
gpus = 1000 / site.p_it_kw
margin = (price - site.cost_per_hour()) * 8760 * site.util * gpus
# closed form of the integral of margin * exp(-r t) from t_fast to t_slow
return margin * (math.exp(-site.r * t_fast) - math.exp(-site.r * t_slow)) / site.r
HYPER = Site("hyperscale", 15000, 12e6, 20, 0.72, 1.25, 0.0905, 600, 0, 0, 0.75)
HOME = Site("home node", 15000, 3e6, 10, 0.72, 1.08, 0.182, 1200, 112.5, 56.25, 0.60)
if __name__ == "__main__":
for s in (HYPER, HOME):
print(f"{s.name:>10}: ${s.cost_per_hour():.3f} per sold GPU-hour")
print(f"home at 75% util: ${replace(HOME, util=0.75).cost_per_hour():.3f}")
print(f"home util needed to match hyperscale: {breakeven_util(HOME, HYPER.cost_per_hour()):.0%}")
for p in (1.64, 1.80, 2.50):
print(f"market ${p:.2f}: home breaks even at {breakeven_util(HOME, p):.0%} util")
v = speed_value_per_mw(HOME, 1.80, 0.5, 3.0)
print(f"value of a 2.5-year head start: ${v / 1e6:.1f}M per MW")The script prints $0.930 and $1.222, the 82% match point, the three market break-evens and the $8.9M head-start value. The model uses a frozen dataclass so that every scenario is an explicit copy made with replace(), which keeps sensitivity runs from mutating the base case. Two sensitivities worth running yourself: double SPAN’s enclosure cost to $6M per MW and the home node rises only to $1.289, while raising O&M to $2,000 per GPU-year takes it to $1.374, so O&M deserves more scrutiny than construction cost.
Which workloads fit: inference yes, frontier training no
A home node is a good inference server and a poor member of a training cluster, and the difference is about four orders of magnitude of bandwidth. SPAN itself positions XFRA for inference, and its launch release cites forecasts that inference will be more than half of all AI workloads by 2030.
Start with inference traffic. A request carries text in and text out. A 100,000-token agent prompt is roughly 400 KB, and streamed output tokens cost tens of bytes each with protocol overhead. A node handling 50 such long-context requests per second moves about 20 MB/s, or 160 Mbps, which fits a residential gigabit fiber line with room to spare. The weights move only when a model is loaded:
t_load = 8·W / B
with W the model size in bytes and B the link speed in bits per second. A 70B model at 8-bit precision is 70 GB and takes 560 seconds over 1 Gbps or 56 seconds over 10 Gbps. A 671B mixture-of-experts model at 8-bit is 671 GB: about 89 minutes over 1 Gbps. That is acceptable for a node that keeps one model resident for days, and unacceptable for a scheduler that wants to swap models every few minutes. Home fleets should be scheduled by model affinity, with each node pinned to a small set of resident models and requests routed to where the weights already live.
Training is a different regime. In synchronous data-parallel training each worker exchanges a full gradient every step. For an 8B model in bf16, a ring all-reduce sends about twice the gradient size, 32 GB per worker per step. Over 1 Gbps that is 256 seconds. Over NVLink at 1.8 TB/s it is about 18 milliseconds. The ratio is 14,400 to 1, which is why no amount of cheap power makes a neighborhood into a pretraining cluster.
The research does narrow that gap. DeepMind’s DiLoCo trains with workers that synchronize only every few hundred inner steps and reports about 500 times less communication while matching synchronous performance on C4 with 8 workers. Syncing every 500 steps cuts the 256-second exchange to about half a second amortized per step. INTELLECT-2 took a more practical route: it kept the gradient updates on trusted, well-connected nodes and pushed only rollout generation, which is inference, out to the swarm, with validators checking the untrusted work. That split is the template. Reinforcement learning, synthetic data generation, evaluation and batch embedding are all inference-shaped, and they are a growing share of what labs spend compute on.
Latency is a smaller advantage than it sounds. Light in fiber covers about 200 km per millisecond, so a node 50 km from the user saves around 15 milliseconds of round trip against a campus 1,500 km away. For a chatbot whose time to first token is hundreds of milliseconds, that barely registers. For voice agents and real-time control loops, where every stage of a pipeline adds latency, it can matter.
There is one more hardware fact to plan around. RTX PRO 6000 Server Edition cards connect over PCIe Gen 5 rather than NVLink, so a model split across all 16 GPUs pays a communication cost inside the node too. The node is best used as several independent replicas of models that fit on one to four GPUs each, or as one large mixture-of-experts deployment where expert parallelism keeps cross-GPU traffic modest.
Where it breaks
The weakest link is the shared transformer at the end of the street. A 200-amp service is 48 kW at 240 volts, and a 12.5 kW node fits inside it with room for an oven and an EV charger. But a neighborhood distribution transformer is shared by several homes and sized for their combined, diversified evening peak. Add a continuous 12.5 kW load to every home and you have multiplied that transformer’s sustained load several times over.
The simulation below makes that concrete. It is vendor-neutral and uses a synthetic household load, so treat the outputs as illustrative; transformer sizes and loading limits vary by utility. A scheduler caps the total node draw so the transformer stays under its rating, and the script reports what share of full-power node-hours the street can deliver.
"""Share a neighborhood transformer between household load and compute nodes."""
import math
import random
NODE_FULL_KW = 12.5 # node at full load
NODE_IDLE_KW = 1.0 # pumps, CPUs, network when curtailed
def house_kw(hour, rng):
"""Synthetic household load: base, evening peak, random big appliances."""
h = hour % 24
base = 1.0 + 2.0 * math.exp(-((h - 19) ** 2) / 6)
spike = rng.choice([0.0] * 8 + [3.5, 7.2, 11.0]) # dryer, EV, oven
return base + spike
def deliverable_fraction(xfmr_kva, homes, nodes, limit=1.0, pf=0.95,
days=60, seed=11):
rng = random.Random(seed)
cap_kw = xfmr_kva * pf * limit
full = 0.0
for step in range(days * 96): # 15-minute steps
hour = step / 4
houses = sum(house_kw(hour, rng) for _ in range(homes))
room = cap_kw - houses - nodes * NODE_IDLE_KW # idle draw is fixed
per_node = max(0.0, min(NODE_FULL_KW - NODE_IDLE_KW, room / max(nodes, 1)))
full += per_node / (NODE_FULL_KW - NODE_IDLE_KW)
return full / (days * 96)
if __name__ == "__main__":
for kva in (50, 75, 167):
row = [f"{deliverable_fraction(kva, homes=8, nodes=n):.2f}" for n in (1, 2, 4, 8)]
print(f"{kva:>4} kVA, 8 homes, nodes 1/2/4/8 ->", " ".join(row))
Transformer, 8 homes 1 node 2 nodes 4 nodes 8 nodes
50 kVA 87% 72% 38% 15%
75 kVA 100% 99% 84% 40%
167 kVA 100% 100% 100% 100%Read the table as a ceiling on utilization before a single customer shows up. On an existing street with a small transformer, a second or third node pushes deliverable hours below the 39% break-even from the cost model. That is why it matters that XFRA starts with new construction: a builder like PulteGroup can size the transformer for the compute load on day one, and the utility studies one subdivision instead of a hundred retrofit requests. The battery in each SPAN install helps at the margin by riding through short appliance spikes, but it cannot fix a transformer that is undersized for a continuous load.
Heat reuse does not scale to GPU nodes. A 12.5 kW node running all day produces 300 kWh of heat. A Heata unit delivers up to 4 kWh of hot water a day, about 167 watts on average, which is 75 times less. A house simply cannot use the heat from an AI node except in deep winter, so the outdoor XFRA cabinet rejects it to the air. Robert Davies of Utah State University estimates that 30% to 40% of homes may be suitable to host a mini data center but that only 2% to 3% could realistically be heated this way, and he warns that such projects tend to be “very light on the cost analysis.” I set the heat credit H to zero in my model for that reason.
Utilization is the commercial risk. The cost model shows a home fleet can survive at around 40% utilization at today’s prices, but those prices rose because supply is short. If campuses catch up in 2028 and rental rates fall the way H100 rates did in 2024 and 2025, distributed nodes with higher field costs will be the first capacity to go idle. Every dollar of O&M is divided by u, so a fleet at 30% utilization pays twice as much per sold hour for its truck rolls as one at 60%.
Security has two halves. Physically, a cabinet holding 16 GPUs worth about $10,000 each is a theft target on a suburban wall. Logically, enterprise customers will ask whether their prompts and weights are safe on hardware a stranger can touch, which pushes operators toward confidential computing, encrypted model storage and remote attestation, all of which add cost the model above does not include.
Neighbors and regulators are the last unknown. Community opposition to large data centers is already intense, with residents near campuses in Texas reporting persistent noise and vibration. SPAN says its units run with minimal noise but has not published a decibel figure. Residential tariffs, zoning codes, HOA rules and insurance policies were not written for a 12.5 kW commercial load on a house, and each will need an answer before a fleet reaches tens of thousands of homes.
Finally, the headline economics are vendor claims. SPAN’s figures for cost per MW, six-times-faster deployment and the 40% to 45% grid utilization hypothesis come from the company. None has been independently audited yet, and the 100-home pilot has not reported results.
How it upends the bottleneck, and what to do this quarter
Home nodes will not replace the campus, and they do not need to. They upend the bottleneck by changing where the marginal gigawatt comes from, and the margin is where compute prices are set.
The capacity math is simple. The US completed 1,010,100 single-family homes in 2025, according to the Census Bureau. If one in ten new homes hosted a 12.5 kW node, that is about 101,000 homes and 1.26 GW of compute capacity a year, with no transmission study and no 160-week substation transformer. LBNL’s projection implies 3.4 to 9.2 GW of new average data center load each year through 2028, so new construction alone could supply something like 14% to 37% of that growth. The comparison is rough, because node nameplate and campus average load are not the same measure, but the order of magnitude is the point: this is a source of supply the size of several large campuses a year, delivered in the window when campuses are stuck in queues. SPAN’s own 2027 target of about 80,000 homes and 1 GW sits inside that range.
What changes is the shape of the supply curve. Today a scarcity rent flows to whoever already holds energized capacity, which is why rental rates rose about 40% in six months. A supply source that can add capacity in six months rather than three to seven years competes that rent away faster, and it moves part of the ownership of AI infrastructure toward homebuilders, panel makers, utilities and the households that host them.
For ML and platform engineers, the work this quarter is to make inference stacks fleet-agnostic. Route by model affinity so weights stay resident, design for nodes that can be curtailed with 15 minutes’ notice, and split RL and synthetic data pipelines the way INTELLECT-2 did, with gradient updates in trusted clusters and generation anywhere. A serving stack that tolerates heterogeneous, interruptible capacity will be able to buy the cheapest hours on whichever grid has them.
For operators and investors evaluating a distributed compute deal, diligence three numbers before the headline cost per MW: realized O&M per GPU-year, contracted utilization, and the transformer headroom per street. The cost model above shows that O&M and utilization move the result more than construction does, and the transformer simulation shows that headroom caps utilization before the market does. Ask for pilot data on all three, and run the speed-value equation with your own estimate of when competing campus capacity actually energizes.
For utilities and regulators, the useful move is a tariff and interconnection path for flexible, curtailable compute load at the distribution level, priced to reward nodes that back off at the evening peak. Without one, every deployment is a one-off negotiation. With one, the distribution grid’s off-peak headroom becomes a product.
I would push back on anyone, including me, who claims home data centers will make the power problem go away. They are a faster, smaller and more operationally demanding way to add inference capacity, and they win only while compute is scarce and campuses are slow. Both conditions hold today. The pilots running over the next year will show whether the field costs hold up as well.
About the Author
Brian James Curry is a Kansas City AI researcher, applied scientist and data scientist working in machine learning, causal inference and healthcare intelligence. He is the founder of Vector1 Labs, an independent applied AI research lab, where his Causal Care series applies causal machine learning to health system decisions: which programs work, for whom, and what they are worth. He also founded KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and has grown to more than 1,500 members.
He has more than twenty years of experience leading analytics, data science and machine learning at Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source research tools on GitHub include PyCausalSim (causal discovery through simulation), Papilon (marketing mix modeling and complex systems) and tincast (forecasting for AI-critical commodities), along with MeaningFlow and Daedalus. A Kansas State University alum from a small Kansas farming town, he writes about agentic AI, causal inference, healthcare AI and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.
Sources
- SPAN, “SPAN Announces XFRA, a Distributed Data Center Solution,” Business Wire, April 2026. Link
- Latitude Media, “Span to launch distributed AI data centers for edge compute,” 2026. Link
- DatacenterDynamics, “Electrical panel company SPAN launches XFRA distributed data center offering,” 2026. Link
- TechRadar Pro, “This US start-up wants to pay you to host a data center on the side of your home,” 2026. Link
- Fortune, “Startups are installing tiny data centers in people’s homes,” 2026. Link
- Startup Fortune, “Startups are paying homeowners to host AI compute nodes,” July 2026. Link
- Centrica, “British Gas partners with Heata on trial to reuse waste heat from data processing,” 2025. Link
- The Register, “Struggling to heat your home? How about 500 Raspberry Pi units?” October 2025. Link
- TechSpot, “Couple heats their home using garden shed data centre,” 2025. Link
- Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, as summarized by Mercom. Link
- IEA, Energy and AI, as summarized by the American Public Power Association, 2025. Link
- LBNL, Queued Up, as summarized by the American Public Power Association, 2026. Link
- Data Center Knowledge, “Why AI data center projects face years of delays after approval,” 2026. Link
- pv magazine USA, “U.S. transformer market faces severe supply constraints as lead times extend to four years,” May 2026. Link
- US Energy Information Administration, Short-Term Energy Outlook, Table 7c, October 2026. Link
- Utility Dive, “EIA: data centers and residential electricity rates,” 2026. Link
- Thunder Compute, “NVIDIA RTX PRO 6000 Blackwell Pricing,” October 2026. Link
- NVIDIA, RTX PRO 6000 Blackwell datasheet, via Ingram Micro. Link
- Crypto Briefing, “Nvidia H100 GPU rental costs surge,” citing SemiAnalysis, August 2026. Link
- CORE Advisors, “AI Data Center Development Costs in 2026,” citing JLL and others, June 2026. Link
- TechRadar Pro, Satya Nadella on chips he “can’t plug in,” November 2025. Link
- Douillard et al., “DiLoCo: Distributed Low-Communication Training of Language Models,” arXiv 2311.08105. Link
- Prime Intellect, “INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning,” arXiv 2505.07291. Link
- US Census Bureau, New Residential Construction, December 2025. Link
- Sify, “You can now host an Nvidia data centre in your home,” 2026. Link
