​

Malibu Litepaper

The world’s biggest AI cloud already exists. Malibu turns it on.

​

The premise

Two hundred million Apple Silicon Macs have shipped since 2020 (methodology). Together they represent the world’s largest idle AI compute network. Each machine delivers roughly ten times the per-watt efficiency of an RTX 5090 on Proof-of-Model-Walk memory-hard weight walks (Phase 1 benchmark, M5 Air, July 2026 — not inference TPS). Fleet idle-time figures are planning assumptions, not audited telemetry — see Benchmarks & methodology. Malibu turns every Apple Silicon Mac into part of the world’s largest AI cloud. Every Mac is a potential provider. Every provider earns from day one. Every dollar of buyer revenue is protocol-converted into demand for the network’s asset. Malibu doesn’t build new AI infrastructure. It activates the infrastructure already sitting on millions of desks.

​

The opportunity

AI inference is a fifty-billion-dollar market growing three-to-five times a year. Three companies own most of it. Open-source models — Llama, Qwen, gpt-oss — have caught up on quality for many workloads. Supply hasn’t. Buyers pay four to twenty times what the underlying compute actually costs. New datacenter capacity takes billions of dollars and years to permit. Meanwhile hundreds of millions of AI-capable Macs sit on desks running at four percent utilization while somebody’s on a Zoom call. Every prior decentralized AI network — Bittensor, Akash, Gensyn, Prime Intellect — hit the same wall: buyers won’t come without providers; providers won’t stay without buyers. A decade of permanent bootstrap. Malibu breaks the wall by paying providers to be ready before buyers arrive.

​

Why now

Five years ago this network wasn’t possible. Today, four preconditions are simultaneously true for the first time:

• Apple Silicon. Unified-memory Macs capable of serving 100B+ parameter open-source models are shipping in consumer hardware.
• Open-source frontier models. Llama, Qwen, gpt-oss now match closed-model quality on many workloads. Verified inference on open weights is finally worth paying for.
• Cheap inference verification. TOPLOC (Prime Intellect, 2025) verifies buyer responses at roughly one percent of serving cost. Cryptographic trust no longer requires trusted hardware.
• Stablecoin payments. USDC on Base makes global micropayments settlement-final in seconds, at sub-cent cost.

Malibu is not early to any of these. It is the network that combines all four.

​

What we’ve shipped

Malibu is a live coordinated inference network today — distributed Apple Silicon providers, centralized routing and billing. Real coordinator, real providers, real inference requests, real settlement on the marketplace rail.

​

Inference marketplace — Live

• Coordinator service. Provider admission, real-time request routing, per-request billing ledger, epoch settlement. Deployed with monitoring, rollback procedures, and nightly reconcile automation.
• Gateway. OpenAI-compatible /v1/chat/completions, /v1/models, /v1/usage, /v1/rate-card. Point any OpenAI SDK client at one endpoint and it works — Cursor, Aider, Claude Code SDK, curl. GitHub OAuth for buyer signup. Quota reservation. Signed session receipts.
• Provider onboarding. Signed public installer auto-detects the Mac’s memory tier, downloads the right MLX model weights, registers with the coordinator, and starts serving. Zero-config from install to first paid request. launchd watchdog for uptime.
• End-to-end serving verified. Real requests flow through the network today: buyer → gateway → coordinator → Apple Silicon provider → response → signed receipt → settlement. The path from request to payment is closed and running.
• Multi-turn tool calling. Standardized tool_calls[] emission and second-turn role: "tool" acceptance for Qwen and Llama-3.3 templates. Cline drop-in works today.
• Structured output. response_format: json_schema with the OpenAI strict-mode subset.
• Sticky conversations. Stable conversation_id routes back to the same provider; cached_prompt_tokens reported in the OpenAI usage object. Prefix-cache metering shipped.
• Signed inference receipts. Every response carries a v0.3 nine-field receipt signed with the provider’s Ed25519 key. Verify with the open-source malibu-verify CLI.
• Verified model settlement. Coordinator receipt verdicts drive gateway buyer debit/refund decisions. Missing or unverifiable receipts settle with zero provider credits.
• Rate card & billing. Per-model pricing served live at /v1/rate-card. Credit-based metering, 90% default provider share today, weekly USDC payout automation with min-payout thresholds.
• Model catalog. Qwen2.5 (7B), Qwen3.5 (35B), Llama-3.2 (3B), Llama-3.3 (70B), Gemma families — MLX-native, memory-tier-matched to provider hardware at install time.
• Release cadence. Signed installer published on GitHub Releases. Full operational runbook: deployment, rollback, monitoring, nightly reconcile, weekly settlement.

​

Token protocol — Planned

• PoMW mining substrate. The Metal port of Proof-of-Model-Walk is public and reproducible on any Apple Silicon Mac shipped since 2020. Two hundred verification vectors passed byte-exact against a Rust reference on real Qwen3-1.7B GGUF weights. Phase 1 benchmarks on an M5 Air: 2.14 million walks per second at 4.7 watts on the walk workload.
• v0 mining launch: on-chain emission, burn-and-mint fee split, and token-native settlement mechanics.
• v1 marketplace: TOPLOC per-request verification (Prime Intellect lineage) on buyer-paid inference.

This is a launch document for a network that already runs inference — and a protocol layer still activating.

​

What Malibu lets you do

Imagine opening your Mac in the morning and realizing it earned money while you slept. No fans screaming. No datacenter. No rack space. No electricity bill worth noticing. Just your laptop, quietly serving AI requests to buyers who need them, all day. Your Mac is no longer just a laptop. It’s an income-producing asset — not a hobby, not a lottery ticket, an actual asset. A single Mac Studio Ultra already serves frontier open models — gpt-oss-120b, Llama 3.3 70B — from unified memory, no datacenter. The models that dominate industry inference spend are larger still, and reaching them is Malibu’s growth vector: pooled serving will link several Macs over a local interconnect to run models beyond any single machine. And the bigger the model, the more a verifiable receipt is worth — when you pay frontier rates, an attestation is the only way to know the endpoint ran the weights you paid for, not a cheaper stand-in. (Pooled serving is on the roadmap, not live.) Under Year-5 base-case assumptions (illustrative, not guaranteed — see Provider economics):

• Mac Studio Ultra: roughly $10,000 to$13,000 per year at $MALIBU =$2. Around $25,000+ at$5. Hardware payback in under twelve months. Even at a bearish $0.50, payback is under twenty-four.
• M4 Max: roughly **$5,100 per year** at$2. Around $12,000 at$5. Payback in under a year.
• M4 Pro: roughly **$3,400 per year** at$2.
• Air: roughly **$1,050 per year** at$2.

Everything after payback is compound income on hardware you were going to buy anyway. Fanless Macs throttle under sustained load — apply a duty-cycle discount (Provider economics). Cooperative scheduling keeps the Mac yours. Mining runs in the background; serving interrupts for milliseconds at a time. It’s just also, quietly, an inference-as-a-service business. For the first time, the compute you already own can pay you back.

​

What buyers get

Companies are massively overpaying for AI inference. Malibu targets 20 to 40 percent savings vs hyperscaler pricing on comparable OSS workloads — methodology and how to verify against the live rate card. Every response carries a signed receipt you can verify offline. The catalog is the OSS frontier: gpt-oss-20b, gpt-oss-120b, Gemma 4 26B A4B, Nemotron, Qwen3 families, Qwen3-Coder — available when pool hardware is warm (Model catalog). Settled in USDC on Base via drop-in OpenAI-compatible endpoints. Why buyers switch:

• Cheaper. Consumer-hardware inference at a fraction of hyperscaler cost.
• Receipt-verified. Signed binding of (prompt, output, model, provider). Per-request forward-pass attestation via TOPLOC ships at v1 — not live on buyer traffic today.
• OSS-native. Best model per task. No vendor deprecating your API next quarter.
• Cooperative trust. Not confidential compute — prompts route through providers as plaintext. See Security & trust model and Threat model.
• No lock-in. Open protocol on Base. Change providers without changing code.

Malibu is designed to be adopted without changing anything on the buyer side except the endpoint.

​

How it works

Every provider runs one binary. At mining launch it performs Proof-of-Model-Walk — a byte-exact 256-step walk through the canonical open-source model’s weights, terminated by a cryptographic commitment. Deterministic. Verifiable in microseconds. The walk hashes over the same weight file the network sells inference for, using the same Metal integer pipeline. Every mined block is a receipt that a specific Mac exercised the memory substrate inference depends on. When a buyer request arrives at v1, the miner yields in ~50 milliseconds — weights already resident, pipeline already running, file already the file the buyer is paying for. Mining state is serving state. That equivalence is the architectural bet the whole network is built on. Verification is cheap. PoMW verifies in microseconds. TOPLOC (Prime Intellect, 2025) will verify buyer responses at about one percent of serving cost at v1.

​

The token flywheel

Every dollar entering Malibu permanently strengthens the network. Providers get paid. Reserves grow. Tokens are bought and burned automatically. When a buyer pays $1 for inference under token economics:

• 70¢ goes to the provider who served the request.
• 12¢ grows a protocol-owned reserves floor — held in USDC, on-chain, backing the token’s minimum value.
• 18¢ automatically buys $MALIBU on the open market and destroys it.

No oracle. No governance vote. No token-holder decision. Every transaction is protocol-enforced $MALIBU demand and protocol-owned reserve growth. The token, minted through work, is burned through use. Provider issuance runs on a deterministic emission curve — front-loaded to bootstrap the fleet, tapering as buyer USDC replaces it.

​

Why Malibu wins

1. Apple already shipped the hardware. Two hundred million Apple Silicon Macs. Strong per-watt PoMW efficiency on consumer silicon. Deployed, paid for, and idle.
2. PoMW prepares the fleet before buyers arrive. Providers earn from day one for verifiable useful work. The cold-start wall dissolves.
3. Receipts today; TOPLOC at v1. Signed receipts bind every response now. TOPLOC adds per-request forward-pass attestation when the marketplace activates.
4. Every transaction strengthens the token. Buyer USDC automatically buys and burns $MALIBU; reserves grow. Network activity IS token demand.
5. The network compounds. Every new provider increases coverage. Better coverage attracts more buyers. Buyer revenue increases provider earnings. Higher earnings attract more providers. Every new participant makes the next one more valuable.
6. Nobody else has a reason to build this. Apple sells hardware. OpenAI sells closed models. Hyperscalers own the datacenters they need to keep utilized. The network that turns consumer Macs into AI infrastructure has to come from outside all three, and it has to be a token network — because only a token network can pay providers to be ready before buyers arrive. The edge is hardware Malibu doesn’t own, and a mechanism its competitors have no reason to build.

​

Join

The world’s biggest AI cloud already exists. Malibu turns it on. Malibu launches with hundreds of providers. It scales to tens of thousands. The economics work best for the ones who show up first: the emission curve pays early miners the most, and mining reputation transfers to serving-revenue priority when v1 activates. Being early isn’t a marketing pitch — it’s the structural incentive the network is designed around. Apple built the hardware. Malibu gives it a job.