The Local Inference Bet

The Thesis

A server in every house is a bet that you can walk around all four walls at once.

On 11 August 2026, Igor Babuschkin’s River AI raised $1.1 billion to build home and small-business servers that run AI locally. The pitch is legible: intelligence you own, in a box you can unplug. It lands under a stronger prediction — that within five to ten years every household runs a local AI server for every workload under the roof, digital or robotic, as ordinary as a router, with a vault so household data never leaves.

This site tracks four binding inputs: the money loop, electricity, memory, and the fabs underneath both. The household thesis is not a fifth wall. It is the claim that you can route around all four by moving the work into houses.

Cost structure says otherwise at every layer. A house buys electricity at retail and silicon at retail, then runs the box at a duty cycle that would end a data-centre career. The walls do not disappear when the work moves home; the household meets each on worse terms, without a balance sheet to negotiate with.

So the useful question is not whether local inference is cheaper. On variable household load it is not. The question is which constraint would have to break for that gap to stop deciding the outcome. This page keeps the list and marks what the record has settled. It does not forecast where any of these stocks go.

The Evidence

Start with what the record is not doing. The screen covers every unique TEXXR article record dated 1 January 2024 through 19 August 2026, matched case-insensitively on headline and TEXXR summary. The local series requires one of on-device, on device, local model, local LLM, run* locally, edge AI, NPU, AI PC or Copilot+ PC, with word boundaries on the last three. The datacentre series requires data center or datacenter. Each record counts once per series per quarter. Both are normalised against total records that quarter, because corpus intake rose from about 3,300 a quarter through 2025 to about 5,500 in 2026, and a raw count would read that as a trend.

In 2024 Q1 the two series were level: 23 records each, 6.58 per thousand each. In 2026 Q2 the local series held 21 records against the datacentre series’ 256 — 3.78 per thousand against 46.13. Across those nine intervals the datacentre series multiplied 7.01x normalised. The local series came to 0.57x of where it started.

The local series is not a rising line with a slow start. It peaked at 58 records in 2024 Q2, on a six-week cluster of Apple Intelligence, Copilot+ PCs and Snapdragon X Elite, and has run between 9 and 22 a quarter since. This is a bet against the visible trend, which is a respectable thing to be. It is not a description of one underway.

Now the nine conditions.

Permission, and whether local resistance pushes load onto retail grids. Live, and documented on the power wall — a first-in-the-nation New York moratorium, 833 opposition groups, roughly $130 billion of projects blocked or delayed in a quarter. But it runs the wrong way for the household. The anger is about bills: PJM power averaged $136.53/MWh in 2026 Q1, up 76% year over year, and its latest auction is reckoned to add $6.3 billion to bills across 13 states through 2029. A home server plugs into the same distribution system whose charges caused the fight, at the residential tariff.

Latency, and robots that cannot wait for a round trip. The record does not carry it. Nvidia’s Cosmos 3 Edge targets robots navigating in real time, AMD’s Ryzen AI Embedded parts are going into humanoid robots, and Arm has split out a Physical AI unit beside separate Cloud and Edge divisions. Each puts compute in the robot — a different product from a house server the robot talks to. Nothing in the corpus quantifies a home-robotics latency budget. Unsettled, and the strongest of the nine on its merits.

A meaningful chip and memory surplus. Refuted outright, and this is the load-bearing failure. NAND contract prices are up more than 600% since end-September 2025 and DRAM nearly 400%. Global DRAM supply is expected to meet only 60% of demand through 2027. It has reached the shelf: the 256GB iPhone 18 Pro carries a bill of materials 38% higher than its predecessor with memory at 34% of it, Apple raised Mac and iPad prices 15–25% on component costs it says have never moved this fast, and smartphone shipments are set to fall 12.9% in 2026, the largest decline on record. The sharpest datum sits at the top of the stack: Nvidia is weighing lower-memory Rubin Ultra parts because it may not secure enough HBM for its own flagship. A household box is a claim on the scarcest input in the economy, held at the lowest utilisation in it. The China-breakthrough branch has not landed: a reverse-engineered EUV prototype in a Shenzhen lab, a domestic DUV tool in testing that tops out near 7nm.

Power as a small share of cost per token. The corpus cannot price the household side, and the page should say so: no residential retail rate, no generation-versus-delivery split, no residential-versus-utility solar cost per watt. What it carries is the shape of the curve. The power book walks TeraWulf’s twenty-year Anthropic lease to roughly $271 per contracted MWh against a ninety-day deal near $5,000, with typical nodes clearing $20–60 most hours. That eighteenfold spread is the market quoting duration, and a household buys at the short end of it, at retail, with no anchor-tenant credit. Against all of it, Sam Altman’s 0.34 watt-hours a query cuts the condition down: if the electricity per query is already trivial, a cheaper electron cannot be what moves a household.

Regulatory and liability cost loaded onto cloud tokens. Supported, and early. Insurers have asked regulators to let them exclude AI chatbot liabilities; QBE and Beazley are moving to cap payouts on AI losses. New York’s RAISE Act gates on revenue — $500 million and above must publish safety protocols and disclose incidents inside 72 hours. Note which way that cuts: it gates the provider, and a household running its own box sits outside it. The cleanest mechanism here — and the one an incumbent can price into a subscription rather than lose a customer over.

Licence terms that push large users to self-generate. The record does not carry the specific claim: no licence in the corpus ties a revenue threshold to Kimi or a comparable release. It carries drift instead — Moonshot moved from a plain open-source K2 to a “modified MIT” K2.6 and K2.7 to a bespoke Kimi K3 License, while DeepSeek stayed on MIT and Qwen and Mistral on Apache 2.0. Directionally live, specifically unverified — and the revenue-gated obligation that does exist in the record is regulatory, not contractual.

A collapse in consumer trust. Refuted on revealed preference. A federal judge ordered OpenAI to produce 20 million chat logs in December 2025; through that, ChatGPT became the fastest app to a billion monthly users. Surveyed Europeans distrust US tech firms with personal data at 84%, and nothing in the corpus shows that converting into switching or a premium paid. The industry’s answer to distrust has been attestation, not relocation: Apple’s Private Cloud Compute invites outside inspection of its server code, and Google shipped Private AI Compute behind it. Most telling, Apple — the privacy maximalist with the best on-device silicon — asked Google to investigate hosting Siri servers in Google data centres under Apple’s privacy bar. The firm best placed to go local went further into the cloud.

Open models banned and driven underground. Live and unresolved. The US framework as drafted excludes open models; the White House is expected to expand it to open models once they reach frontier capability, and parts of the administration are reviving de facto bans on foreign ones. That is hard to do quietly when Chinese models are already nearly 60% of token usage by US companies on OpenRouter. And the capability that would make a ban bite is arriving: open weights now trail the closed frontier by four to seven months on cyber evaluations, narrowed from six to ten.

Consolidation holding cloud margins above 90%. The premise fails, and its failure closes the argument rather than weakening it. Anthropic projected a 40% gross margin on 2025 AI sales, revised down from about 50% because inference ran 23% over plan. Oracle’s Nvidia-rental line turned roughly $900 million of revenue into $125 million of gross profit, a 14% margin against a corporate rate near 70%. At the consumer tier the subsidy runs the other way: $200-a-month plans deliver up to $8,000 and $14,000 of tokens a month at API rates. There is no 90% cushion to eat into. The household is being asked to beat a price already sold below cost. Pricing power exists — DeepSeek raised V4 prices in August, taking Flash output from $0.28 to $1.32 per million at peak, months after cutting it 75% — but that is one provider repricing into demand, not an oligopoly holding a floor. Nor is the underlying task cost falling cleanly: the price of a token has dropped while developer bills have risen, because reasoning models spend far more tokens per job. A household box pays that tax at a far worse duty cycle.

Two reversals deserve stating rather than burying. In March 2025 Amazon told Echo owners they could no longer process requests locally, because the new generative features needed the cloud — from the company that had built a dedicated on-device speech chip five years earlier. In 2024 IDC found only about 3% of PCs shipped would clear Microsoft’s AI PC threshold, with some large app makers rebuffing a push to move work on-device. Both come from the parties best positioned to make local work.

Meanwhile the cloud side compounds. Google went from 9.7 trillion tokens a month to 480 trillion to 3.2 quadrillion in two years. Frontier models grew from Llama 3’s 405 billion parameters to Kimi K3’s 2 to 3 trillion. Alibaba Cloud’s pooling claims it cut required H20s by 82% serving dozens of models at once — a batching gain that exists only where many tenants share one machine, and that a single household cannot reproduce.

The Companies

There is no basket for this trend, and the absence is the finding. The registry has no consumer or edge-silicon group; the nearest are memory and accelerators, both of which sell into racks. A reader looking for a listed pure-play on household inference will not find one, because the credible builder of it raised private money this month.

The hardware exists and is priced for developers. Nvidia sells the DGX Spark at $3,999; Microsoft built a Surface RTX Spark Dev Box around the same Arm part with 128GB of unified memory. Apple’s most capable on-device model needs an iPhone 17 Pro or better with 12GB of RAM, and Tim Cook has said Mac Studio and mini supply, bought heavily for agent work, may take months to balance. Ollama, the tool most people use to run open weights locally, raised a $65 million Series B against River AI’s $1.1 billion. That ratio is the market’s read on where local inference is a product and where it is a workshop.

The names with a real claim here get paid either way. $MU and $SNDK, with SK Hynix and Samsung beside them on the memory wall, supply the bit a home server needs more of per unit of useful work than a rack does, because it cannot amortise a model across tenants. Every household box shipped competes for the same wafers as an accelerator — which is why memory expresses this argument more cleanly than edge silicon does. $NVDA sits on both sides and has shown which one it believes: a $3,999 developer appliance on one, rack-scale inference on the other. $INTC and $AMD carry the AI PC and embedded-robotics ends and $ARM the licensing layer under both; none has a dossier here yet.

The cloud path runs through names already covered — $GOOGL, $MSFT and $AMZN at the platform layer, $CRWV, $NBIS and $ORCL renting capacity, with the capex loop financing it. If the household thesis lands, that is where revenue leaves from. Nothing in the record suggests it is leaving.

The Lenses

The Helmer lens asks which of the seven powers a household server could hold, and the answer is close to none. Scale economies run the wrong way by construction: unit costs fall as volume rises, so a rack serving thousands of tenants prices where a single-tenant box loses money on every token. There is no cornered resource — the household is at the back of the memory queue. There is no network economy in a machine designed so that nothing leaves the building; the vault that makes it appealing is the feature that denies it one. What remains is counter-positioning, which is exactly River AI’s stated pitch: trainable, and not controlled by a big company — a thing no hyperscaler can offer without damaging its own business. Counter-positioning is a real power. It is also the one that most often turns out to be a niche rather than a market.

The Thaler lens explains why the survey data and the adoption data disagree. Mental accounting files a $4,000 appliance and a $20 subscription in different ledgers, and the subscription wins even where three-year arithmetic does not favour it. Inertia keeps a household on a default it has already configured. Stated preference about privacy is nearly free to express; revealed preference is a billion monthly users through a court-ordered disclosure of twenty million chat logs. The lens does not say people are wrong to want privacy. It says a claim resting on consumers acting on a preference they have not yet acted on is the weakest kind of forecast — and this one rests on it twice, for the purchase and for the switch.

What Moved

The DeadRisk market layer, reconciled against the point-in-time price archive, records 18 August 2026 as a broad down day across this cast. Against the 17 August close, $SNDK fell 9.01%, $MU 7.02%, $ARM 6.67%, $INTC 6.58% and $AMD 4.27%, with $NVDA down 2.34%. Two held: Apple rose 1.45% and $MSFT 0.27%. Non-US names in this cast sit outside the point-in-time archive and are not quoted here.

Over the 21 sessions through that close the split is sharper than the day. Memory led — $SNDK up 16.88% and $MU up 8.70% — while every name that would have to build the household box lagged or fell: $AMD down 3.81%, $INTC down 0.39%, $ARM down 6.04%, Apple down 5.07%. The record establishes the returns and the divergence. It does not establish a cause.

Sources