Date: June 2026 · Not financial advice — research to inform your own decisions.
Real-world earnings data (not marketing) makes this clear: the cheapest viable entry is high-VRAM NVIDIA hardware you already own or buy used — and "lots of small compute" (Pis, Mac Minis) is not a viable route at all. Worse for the small-device idea: even centralised Mac compute fetches almost nothing — Scaleway rents an entire dedicated Mac Mini M4, in a datacentre with SLAs, for €0.22/hr. A home Mac on a trustless network would price below that ceiling, and today no live DePIN pays for Apple Silicon anyway (Nosana's "Whirlpool" Apple-Silicon support was roadmapped for H1 2026; I found no evidence it has shipped — unverified). Raspberry Pis remain a hard zero: no CUDA GPU, ineligible for the paying workloads.
Real host data across Vast.ai / Salad / RunPod-type marketplaces, 2026:
| Hardware | Gross/month (good uptime) | Source quality |
|---|---|---|
| RTX 3080 (10GB) | ~$25–70 | Aggregated host guide |
| RTX 3090 / 4080 (24/16GB) | ~$50–130 | Aggregated host guide |
| RTX 4090 (24GB) | ~$90–250 | Aggregated host guide; Salad's own marketing says "up to $150/mo" |
| 4× RTX 5090 (real operator) | $460/mo total (~$115/card) | Documented operator result, May 2026 |
Cross-check from the buyer side: SaladCloud rents a 32GB RTX 5090 to buyers at $0.25/hr (~$180/mo continuous). That's the revenue ceiling per card before the platform's cut — which is why per-card host earnings land where they do. (Ignore the "$500–1,000/mo per 4090" claims circulating from platform marketing; the documented operator data above contradicts them.)
High VRAM matters more than raw speed — AI workloads are memory-hungry, which is why a used 24GB 3090 out-earns newer 12–16GB cards.
Assumptions: ~30–50% utilisation, system draw ~550W under load / ~100W idle → ~£40–50/mo power for a single-GPU rig running 24/7.
| Entry option | Upfront | Net £/month (realistic mid) | Payback | Verdict |
|---|---|---|---|---|
| Hardware you already own (e.g. existing 4090-class render kit, idle between jobs) | £0 | ~£40–110 | Immediate | The only clearly rational entry. Pure idle-cost recovery. |
| Used RTX 3090 24GB (~£650) | £650 | ~£15–60 (mid ~£35) | ~18–36 months | Marginal. Best £/VRAM entry if buying. |
| New RTX 4090 (~£1,600) | £1,600 | ~£30–145 (mid ~£85) | ~19 months (11 best-case → never at low utilisation) | Akash's "13-month payback" assumes 50% utilisation; sector data says expect less. |
| 4× RTX 5090 rig (~£8,000+) | £8,000+ | ~£250–350 | ~2.5+ years | Documented real case. Capital-inefficient. |
| Raspberry Pi cluster (£700–900) | £700–900 | ~£0 | Never | Ineligible for paying workloads. |
| Mac Minis | sunk/owned | ~£0 today | n/a | No live network pays for Apple Silicon; €0.22/hr centralised ceiling means it'd be pennies even if one did. Watch Nosana Whirlpool. |
Timeline reality: payback windows of 18–36 months run concurrently with GPU depreciation over the same period, and earnings are partly token-denominated (volatile). The strategy that maximises whatever ROI exists: multi-platform stacking — one rig listed on Salad + Vast.ai + Nosana/Akash simultaneously with auto-switching (hybrid setups pause one workload when a better-paying job lands), which is exactly the orchestration-layer play identified earlier.
Run the two routes against each other: spending new money on GPUs to farm DePIN income is a worse investment than simply buying the tokens. A £1,600 4090 returning ~£85/mo net is ~6% monthly yield before depreciation, downtime, ops effort and token volatility — and it caps your upside at the rental rate. Token exposure (Part 2) captures the same sector thesis with zero ops and uncapped (also uncapped-downside) exposure. So:
UNI/CRV/LINK-style: token attached to a product with real-world usage and income. The sharp version of that test — the one UNI itself famously failed for years (real product, billions in fees, but no fee switch: the token captured none of it) — is:
Scoring the DePIN set against all four:
| Token | Real product & users | Real income | Token captures it? | Thesis verdict |
|---|---|---|---|---|
| AKT | ✔ audited leases, 40–50% GPU utilisation | ✔ ~$0.8–1M/qtr, on-chain | ✔✔ Strongest mechanism in the set: 1–2% take-rate on every lease → community pool, plus BME (usage market-buys & burns AKT; vault removes float) | Cleanest thesis fit. Small income, but the pipe from usage → token is direct and audited. The "LINK-like" pick. |
| RENDER | ✔ 1.5M frames/mo, AI 35–40% of volume | ◐ revenue undisclosed, but usage on-chain via burns | ✔ Burn capture: ~95% of every job's spend is burned; burns +279% YoY; not yet net-deflationary (~120K burned vs ~500K minted/mo) — crossover is the event | Good fit, pre-inflection. Value accrual = scarcity, not income share. You can verify its demand side personally — it's your industry. |
| TAO | ✔ Chutes 400k users; $43M Q1 subnet revenue | ✔ largest real revenue in the set | ◐ Indirect: income flows to miners/validators/subnet owners, not TAO holders; TAO captures via staking demand (>70% locked), subnet-registration burns, and being the base pair for all subnet tokens | Partial fit, biggest scale. More "reserve asset of an economy" than fee-capture token. Halving + ETF path is the kicker, but it's not a UNI/CRV analogue. |
| IO | ◐ revenue growing ($5.7M/qtr) but verified supply shrinking | ✔ | ◐ fee/burn exists; IDE (Q2 2026) would strengthen it — if executed | Fails on trust + 2025–27 unlock overhang. Event-driven only. |
| NOS | ◐ real but small (Sogni, ~2k nodes) | small | ◐ fees/staking exist but thin | Fails on scale: $22M mcap, 42% whale concentration. Lottery ticket, not a thesis holding. |
For a long-play, usable-product-token allocation in this sector, the analysis lands on:
Core: AKT + RENDER — they're the two that pass the capture test mechanically:
Satellite (optional, different thesis): TAO — not a fee-capture token, but the only one with institutional-scale real revenue and the post-halving/70%-staked supply structure. Hold it as "the sector's reserve asset," sized smaller, with the ETF decision as the swing event. Skip if you want to stay strictly inside the UNI/CRV/LINK frame.
Avoid for this thesis: IO, NOS — IO until the IDE tokenomics actually ship and unlocks digest; NOS at this liquidity/concentration profile.
Position discipline that the mechanics dictate: these all have usage-linked buy/burn pressure, so the leading indicators genuinely lead — Messari lease revenue (AKT), burn-vs-mint (RENDER), taostats subnet revenue + ETF dates (TAO). Set a review cadence around those releases rather than the chart. And given DCA logic applies doubly in a sector down 70–95% from highs with improving fundamentals: staged entry beats lump-sum, and pre-decide your fiat-conversion rule.
The clean structure for you specifically:
That loop — earn a little, learn a lot, position the capital on what you learn — is the most efficient low-cost entry available.
Supply-side / earnings
Token thesis
Caveats: earnings figures are gross ranges from host-reported data at mid-2026 rates and shift with demand; UK power modelled at £0.27/kWh; payback windows ignore token volatility and GPU resale value (which partially offsets depreciation). Nosana Apple-Silicon support status is unverified-as-shipped. Token analysis is point-in-time; nothing here is investment advice.