🤫 Puppy One — the Summer 2026 catalog
A supercomputer on every desk that can afford it, by the second.
From the phone already on your monthly bill to AI factories in orbit — one honest ladder, priced by the watt and by the second, with consent built in at every rung. One Puppy at every watt.
Every system, one honest scoreboard.
Vendors put their best foot forward against a published formula - not a hundred opinions. Four axes, ranked openly.
Sustained useful output on reference workloads for the rung class.
The per-watt outcome. Efficiency is the product, not a footnote.
Capex over 60 months plus power - or dollars per hour on cloud rungs.
How long an agentic loop runs unattended. Hours on a phone. Months in a factory.
The Ladder · Ground
It starts in your pocket.
Rungs 1–6: silicon you own. Every dollar is deductible capex under 100% bonus depreciation.
iPhone 17 Pro · Galaxy S26 Ultra · Pixel 10 Pro — a billion NPUs already ship, metered and on autopay.
Jetson Orin Nano Super at $249 · Mac mini M4 at $599 · Strix Halo minis — humanity-scale entry points.
MacBook Pro M5 Max 128GB · Legion Pro 7i 5090 · ZBook Ultra G1a — mobile 70–120B-class inference.
DGX Spark · dual Spark 256GB · Dell Pro Max 18 Plus · 5090 towers — the first rung built to never sleep.
RTX PRO 6000 96GB towers · Threadripper PRO quads — fine-tuning and heavy pipelines at the desk.
DGX Station GB300 with 784GB · 8× DGX B300 — trillion-parameter inference beside the Ferrari.
The Ladder · Sky
H100 median ~$2.30–3.12 · B200 $2.69–16.11 · spot from ~$1 — neoclouds beat hyperscalers on every chip.
Lambda 1-Click 16–2,000+ GPUs · TPU pods · Trainium2 · Core42 UAE sovereign — the factory floor.
The Ladder · Space
Starcloud-2 · Project Suncatcher · Starlink-connected field nodes — identities in space may be many.
The Ladder · Back to Earth
Claude · GPT · Gemini · Grok · DeepSeek — buy the outcome, skip the silicon entirely.
The Puppy 100
Ten rungs. Ten systems each. One handshake.
Every entry is a certified Puppy host with a spaceID. Prices marked ✓ were verified July 7, 2026; ~ is a market band to confirm at order time; ⏳ is upcoming or waitlist.
Showing 100 of 100 systems
| R | System | Band | Loop | Status |
|---|---|---|---|---|
| R1 | iPhone 17 Pro MaxA19 Pro NPU — the default Puppy host | on your bill | hours | ~ |
| R1 | iPhone 17A19 — consent runtime in every pocket | on your bill | hours | ~ |
| R1 | Samsung Galaxy S26 UltraSnapdragon 8 Elite Gen 5 | on your bill | hours | ~ |
| R1 | Google Pixel 10 ProTensor G5 on-device Gemini | on your bill | hours | ~ |
| R1 | OnePlus 15Flagship NPU, unlocked-friendly | on your bill | hours | ~ |
| R1 | iPad Pro (M5)Laptop-class NPU, tablet body | $999+ | hours | ~ |
| R1 | iPad Air (M3)The affordable big-screen Puppy | $599+ | hours | ~ |
| R1 | Galaxy Tab S11 UltraAndroid's big-canvas node | $1.1K+ | hours | ~ |
| R1 | Nothing Phone (3)Design-forward budget NPU | $799 | hours | ~ |
| R1 | Meta Quest 3SSpatial edge inference, worn | $299 | hours | ~ |
| R2 | NVIDIA Jetson Orin Nano Super67 TOPS dev kit — humanity's entry ticket | $249 | hours–days | ✓ |
| R2 | Raspberry Pi 5 16GB + AI HAT+The people's AI computer | ~$200 | hours | ~ |
| R2 | Mac mini (M4)Smallest Mac, serious NPU | $599 | hours–days | ✓ |
| R2 | Mac mini (M4 Pro)64GB unified in a palm | ~$1.4K | days | ~ |
| R2 | Beelink SER9 (Ryzen AI 9 HX 370)Mini-PC NPU workhorse | ~$999 | days | ~ |
| R2 | Minisforum AI X1 Pro96GB-capable Strix Point mini | ~$1.2–1.5K | days | ~ |
| R2 | GMKtec EVO-X2 (Ryzen AI Max+ 395)Strix Halo 128GB in a mini | ~$1.5–2K | days | ~ |
| R2 | ASUS NUC 15 ProEnterprise-manageable small node | ~$800+ | days | ~ |
| R2 | Jetson AGX Orin 64GBRobotics-grade edge brain | ~$2K | days | ~ |
| R2 | HP Z2 Mini G1a (Strix Halo)Workstation mini, 96GB to GPU | ~$1.9K+ | days | ~ |
| R3 | MacBook Pro 16 M5 Max 128GB614GB/s unified — the sovereign inference laptop | ~$5.1K | hours–days | ✓ |
| R3 | MacBook Pro 14 M5 ProThe everyday pro Puppy | from ~$2.2K | hours–days | ~ |
| R3 | Lenovo Legion Pro 7i Gen 10 (RTX 5090)Best $/TFLOP in a laptop — full 175W | $3.8–4.1K | hours–days | ✓ |
| R3 | ASUS ROG Flow Z13 (128GB Strix Halo)Tablet-shaped unified-memory monster | ~$2.8K | hours–days | ~ |
| R3 | HP ZBook Ultra G1a 128GB112GB to the GPU at 3 lbs | ~$4.3K | hours–days | ✓ |
| R3 | Razer Blade 16 (RTX 5090)Thin-and-fast CUDA flagship | ~$4.5K | hours–days | ~ |
| R3 | Dell Pro Max 16 (RTX PRO)ISV-certified mobile workstation | ~$4–7K | hours–days | ~ |
| R3 | HP ZBook Fury G1i 18 (RTX PRO 5000)192GB RAM ceiling, 150W GPU cap | ~$8K+ | days | ~ |
| R3 | MSI Titan 18 HXDesktop-replacement maximalist | ~$5–6K | hours–days | ~ |
| R3 | Framework Laptop 16Modular, repairable, upgradable | ~$2–3K | hours–days | ~ |
| R4 | NVIDIA DGX Spark128GB unified, full CUDA on a desk | $4,699 | days–weeks | ✓ |
| R4 | ASUS Ascent GX10GB10 for less | ~$3K | days–weeks | ~ |
| R4 | AMD Ryzen AI Halo dev box128GB unified, 300B-ready line | $3,999 | days–weeks | ✓ |
| R4 | Dual DGX Spark (linked)256GB pooled — 405B-class at home | ~$9.4K | weeks | ✓ |
| R4 | Framework Desktop 128GBOpen Strix Halo tower | ~$2.4K | days–weeks | ~ |
| R4 | Dell Pro Max 18 Plus (Pro 5000 · 175W)128GB CAMM2 @7200 — the uncapped flagship | ~$8.0–8.6K | days–weeks | ✓ |
| R4 | Mac Studio (M3 Ultra 96GB)819GB/s — capped by the DRAM cycle | $5.5K · 13–14wk | weeks | ~ |
| R4 | Mac Studio (M5 Ultra)The memory king returns | H2 2026 | weeks | ⏳ |
| R4 | RTX 5090 tower (Puget / DIY)21,760 cores, 1.79TB/s | ~$5–7K | weeks | ~ |
| R4 | ThinkStation P3 Ultra (RTX 5090)SFF pro tower | ~$5–6K | days–weeks | ~ |
| R5 | RTX PRO 5000 72GB tower70B unquantized on one card | ~$10–15K | weeks–months | ~ |
| R5 | RTX PRO 6000 Blackwell 96GB towerThe single-card ceiling | ~$15–20K | weeks–months | ~ |
| R5 | Dual RTX PRO 6000 · Threadripper PRO192GB VRAM under the desk | ~$30–45K | months | ~ |
| R5 | HP Z8 Fury G5Quad-GPU classic | ~$20–60K | months | ~ |
| R5 | Lenovo ThinkStation PXDual-socket, quad-GPU | ~$20–60K | months | ~ |
| R5 | Dell Pro Max 7 TowerEnterprise heavy iron | ~$15–50K | months | ~ |
| R5 | Puget Systems quad-GPUBuilt-to-order US workstation | ~$40–60K | months | ~ |
| R5 | Supermicro SYS-551AServer-grade at the desk | ~$25K+ | months | ~ |
| R5 | tinybox pro8-GPU open-stack box | ~$40K | months | ~ |
| R5 | Comino GrandoLiquid-cooled multi-GPU | ~$30–50K | months | ~ |
| R6 | NVIDIA DGX Station GB300784GB unified — trillion-param inference | $80–125K | months | ✓ |
| R6 | HP ZGX Fury AI Station G1nGB300 in HP clothing | ~$100K+ | months | ~ |
| R6 | Dell Pro Max GB300 stationSame silicon, Dell service | ~$100K+ | months | ~ |
| R6 | DGX B300 (8-GPU system)The garage-rack flagship | $300–350K | months | ✓ |
| R6 | Supermicro HGX 8× H200Proven 8-way workhorse | ~$250–300K | months | ~ |
| R6 | Lambda Vector ProTurnkey research rack | ~$100K+ | months | ~ |
| R6 | GB200 NVL4 nodeGrace-Blackwell building block | ~$120K+ | months | ~ |
| R6 | ASUS ESC N8-E118-GPU HGX server | ~$250K+ | months | ~ |
| R6 | Gigabyte G593 (HGX B200)Dense Blackwell 5U | ~$300K | months | ~ |
| R6 | QCT QuantaGrid 8-GPUHyperscale-grade OEM | ~$250K+ | months | ~ |
| R7 | NVIDIA H100 (cloud)47 providers, 23× price spread | $1.38–8/hr · med $2.3–3.1 | unlimited | ✓ |
| R7 | NVIDIA H200 (cloud)141GB — 70B without sharding | $1.00–10.60/hr | unlimited | ✓ |
| R7 | NVIDIA B200 (cloud)192GB, 8TB/s — repriced +24% in March | $2.69–16.11/hr | unlimited | ✓ |
| R7 | NVIDIA B300 (cloud, early)288GB Blackwell Ultra | $4.95–18/hr · spot ~$2.90 | unlimited | ✓ |
| R7 | NVIDIA A100 (cloud)The value workhorse | $1.29–2.50/hr | unlimited | ✓ |
| R7 | AMD MI300X (cloud)192GB HBM alternative | ~$2–4/hr | unlimited | ~ |
| R7 | NVIDIA L40S (cloud)Inference value tier | ~$0.8–1.8/hr | unlimited | ~ |
| R7 | NVIDIA GH200 (cloud)Grace-Hopper superchip | ~$3–6/hr | unlimited | ~ |
| R7 | Consumer 4090/5090 marketplacesVast / Salad — the people's spot market | from ~$0.20/hr | unlimited | ~ |
| R7 | Serverless per-second GPUModal / RunPod — pay only while thinking | usage | unlimited | ~ |
| R8 | Lambda 1-Click Clusters16 to 2,000+ B200s, self-serve | contract | months–years | ✓ |
| R8 | CoreWeave reservedNeocloud at factory scale | contract | months–years | ~ |
| R8 | Nebius AI clustersEU-flavored capacity | contract | months–years | ~ |
| R8 | Google Cloud TPU v6e/v7 podsThe TPU lineage — Manish shipped it | contract | months–years | ~ |
| R8 | AWS Trainium2 UltraServersCustom-silicon economics | contract | months–years | ~ |
| R8 | Azure ND GB200 v6Blackwell on Azure fabric | contract | months–years | ~ |
| R8 | Oracle OCI SuperclusterZettascale marketing, real racks | contract | months–years | ~ |
| R8 | Core42 (G42) sovereign — UAEThe Abu Dhabi lane 🇦🇪 | contract | months–years | ~ |
| R8 | GB200 NVL72 rack72 GPUs, one NVLink domain | ~$3M class | months–years | ~ |
| R8 | AI-factory colocation (Stargate-class)Gigawatt campuses, reference rung | reference | years | ~ |
| R9 | Starcloud-2 · Crusoe on-orbitFirst Blackwell in space — Oct 2026 launch | capacity early 2027 | — | ⏳ |
| R9 | Google Project SuncatcherTPU prototype satellites | early 2027 | — | ⏳ |
| R9 | SpaceX orbital DC constellation1M-satellite FCC filing | filed Jan 2026 | — | ⏳ |
| R9 | Blue Origin TeraWave5,400-sat DC network | announced | — | ⏳ |
| R9 | Axiom Space ISS data centerOrbital nodes to station | 2027 | — | ⏳ |
| R9 | Lonestar lunar backupData vault on the Moon | niche · live | — | ⏳ |
| R9 | Aethero orbital JetsonFirst space Jetson, flown | flown 2025 | — | ⏳ |
| R9 | ESA Space CloudEuropean orbital DC program | study 2026 | — | ⏳ |
| R9 | Starlink Mini + Jetson field nodeEdge anywhere the sky is visible | ~$1K | days | ~ |
| R9 | 🤫 Garage Grid nodespaceID-certified, consent-metered host | spec 2026 | months | ⏳ |
| R10 | Anthropic Claude APIWhere the longest loops live today | $/Mtok | months | ✓ |
| R10 | OpenAI APIGPT + o-series reasoning | $/Mtok | months | ~ |
| R10 | Google Gemini APILong-context native | $/Mtok | months | ~ |
| R10 | xAI Grok APIColossus, rented by the token | $/Mtok | months | ~ |
| R10 | DeepSeek APIOpen-weight frontier economics | $/Mtok | months | ~ |
| R10 | Mistral APIEuropean open-weight lane | $/Mtok | months | ~ |
| R10 | Llama via Together / FireworksOpen models, managed speed | $/Mtok | months | ~ |
| R10 | Groq Cloud (LPU)Deterministic-latency inference | $/Mtok | months | ~ |
| R10 | Cerebras InferenceWafer-scale tokens | $/Mtok | months | ~ |
| R10 | OpenRouterEvery model, one meter | $/Mtok | months | ~ |
You already pay for a supercomputer every month.
Metered billing. Autopay. Universal reach. The rails exist - compute is the eighth utility, and it rides the bills you already trust.
The wedge is in your pocket
Carriers already bundle devices on installments. The SIM is the spaceID root of trust - Rung 1 costs the customer nothing extra.
The Puppy box, $6.90/mo
Your modem rents for $10 a month. A Rung-2 Puppy node rides the same equipment line - private AI at home, on the bill you already autopay.
The garage joins the grid
Rung 4–6 owners pair 100% first-year expensing with consented compute rental and demand-response - silicon that earns while it idles.
Compliance by design: the carrier or ISP is always merchant of record with explicit opt-in - no third-party bill cramming, ever. Consent is the product.
OBBBA · 2026
Expense 100% in year one.
Rungs 2–6 are qualified capex. Rungs 7–10 are opex, deducted as incurred. Every buyer keeps more of what they build with.
Permanent bonus depreciation on hardware acquired after Jan 19, 2025.
2026 Section 179 ceiling - with a $4.09M phase-out threshold.
Identity
One person. Many machines. One handshake.
The eSIM is the hardware root of trust; Starlink direct-to-cell gives every node a SIM presence - on Earth and off it. Every job any machine runs is one PCHP handshake:
The key to the grid
Reserve your place on the ladder.
One reservation holds your place in line and mints your spaceID - your machine's identity on the grid, bound to you. Your silicon, your business.
One is a product of Hushh Technologies Corporation (brand: 🤫 “hussh”), an independent company. One runs on third-party silicon, systems, and cloud; all company names are used solely to describe the platforms on which One software runs. Hushh Technologies is not affiliated with, endorsed by, sponsored by, or partnered with any company named.