Decentralized LLM Inference on Consumer Apple Hardware — Economics & Feasibility

Research date: 2026-09-30. Scenario: individuals rent out Macs overnight (~8 h/night) to serve LLM inference to paying customers. Reference host: M5 Max Mac Studio (2026). Buyers reach hosts via tunnel + central marketplace/API gateway.

Summary

The arbitrage does not work if you buy the Mac for this business — at commodity 2026 token prices, a $5,399 M5 Max Mac Studio (128GB) can gross roughly $0.75–$2.60 per 8-hour night at realistic throughput and 100% utilization, against ~$5.09/night in amortized hardware + electricity cost. Break-even would require >200% utilization: impossible.

The arbitrage is real as a marginal-cost play on already-owned idle Macs: electricity is $0.16/night, so any revenue above that is profit ($18–74/month per machine at realistic utilization). It can also work in premium niches — private/bring-your-own-model inference, dedicated flat-rate nightly nodes, or batch workloads — where buyers pay 2–3× commodity token prices. Three structural facts drive this verdict:

1. Token prices for the model classes that fit in 128GB have commoditized ($0.10–0.60/1M tokens) while the machine costs $5,399.
2. True 2026 frontier open models (1T+ parameters) do not fit in 128GB, so hosts cannot charge frontier prices; they compete in the GPT-4o-class tier.
3. Decode throughput on Apple Silicon is memory-bandwidth-bound and modest for dense 70B models (~13 tok/s); the economics only approach viability with MoE models (few active params) plus a continuous-batching server.

1. Hardware: M5 Max Mac Studio

• Specs/pricing: https://wavect.io/blog/mac-mini-m6-vs-mac-studio-m5-local-ai/ ; https://www.macobserver.com/news/mac-studio-m5-max-vs-m5-ultra-configurations-compared/ ; https://www.macworld.com/article/2973459/2026-mac-studio-m5-release-date-specs-price-rumors.html (all read 2026-09-30)
• 128GB/1TB = $5,399: https://pub.towardsai.net/m5-max-128gb-or-m5-ultra-96gb-the-mac-studio-buying-guide-for-local-llms-in-2026-5b7171d482af (read 2026-09-30); cross-checked live against the Apple.com configurator on 2026-09-30: base $2,499 + $300 (40-core GPU) + $2,000 (128GB) = $4,799 at 512GB SSD; +$600 for 1TB = $5,399. Arithmetic matches exactly. (https://www.apple.com/shop/buy-mac/mac-studio)

Why unified memory + bandwidth are the whole game

LLM decode (generating one token at a time) is memory-bandwidth-bound: each generated token requires streaming the model’s full weight matrix through the processor once. On Apple Silicon the GPU and CPU share one unified pool, so a 40GB model can sit entirely in fast memory with zero PCIe copying (unlike a discrete NVIDIA GPU, where weights must fit in VRAM).

Estimation math: decode tok/s ≈ (memory bandwidth × efficiency) ÷ bytes per token, where bytes per token ≈ model size in bytes (all weights read once per token), adjusted for quantization (Q4 ≈ 0.5 bytes/param + overhead).

Worked examples (M5 Max, 614 GB/s, assuming ~75–85% achieved efficiency):

• 70B dense Q4 (~40GB on disk): 614 ÷ 40 = 15.4 tok/s theoretical → measured 11–13.1 tok/s in published benchmarks (see §3). Theory ✓ matches measurement.
• MoE 30B-A3B Q4 (~3B active params/token, ~17GB): decode moves only active experts + shared layers per token, so effective bytes/token ≈ 4–5GB → 614 ÷ 4.7 ≈ 130 tok/s measured. This is why MoE models with few active parameters are the sweet spot on Apple Silicon: throughput is set by active bytes, not total params.
• KV-cache grows with context length and concurrency and eats into the 128GB budget alongside weights; long-context + many concurrent users can OOM a host even when weights fit.

2. Models: Frontier-class open weights (mid/late 2026)

The 2026 open-weight frontier is Chinese-led; Meta’s Llama 4 has slipped and its frontier effort (Muse Spark) is closed/API-only ( — https://medium.com/@surajkhaitan16/i-benchmarked-the-top-20-llms-of-2026-heres-which-to-use-for-what-f2447efc5954, read 2026-09-30). None of the true frontier open models fit in 128GB — they are 400B–1.6T parameter MoEs. Realistic candidates that fit a 128GB M5 Max:

• Model landscape/quality: https://codersera.com/blog/best-open-source-llm-2026-llama-4-qwen-3-5-deepseek-v4-gemma-4-mistral/ ; https://tech-insider.org/llama-4-vs-qwen-vs-mistral-2026/ ; https://callsphere.ai/blog/llm-comparison-code-generation-open-vs-open-may-2026 (read 2026-09-30)
• “Matches or beats GPT-4o” for 70–80B class: https://github.com/neo-x/llm-playground/blob/HEAD/LLMs%20performance%20analsys.md (read 2026-09-30)
• All four fit in 128GB with room for OS + KV cache (measured: 70B Q4 peaks at 41.2GB; 123B Q4 at 69.2GB — see §3). All are Apache 2.0 = cleanest for commercial hosting (see §6 risks).
• What doesn’t fit: Kimi K3 (top open index 57, but custom license + huge), DeepSeek V4-Pro (1.6T/49B, MIT), Llama 4 Maverick (400B/17B, ~220GB at Q4), Qwen3.5-397B (~200GB at Q4), DeepSeek 671B-class (~360GB at Q4). These need an M5 Ultra 512GB ($12k–18k) or multi-host — outside the consumer-overnight scenario. So hosts compete in the GPT-4o-class tier, not the 2026 frontier tier, and must price against commodity providers accordingly.

3. Throughput: Realistic tokens/sec on M5 Max

Published M5 Max (128GB, 614 GB/s) measurements, all via MLX-family runtimes:

Notes:

• Prefill is the weak spot on Macs: ~350–450 tok/s on a 4K prompt at 7B; a 100K-token prompt takes minutes ( — https://github.com/t0rsion/leone/blob/HEAD/research/07-consumer-hardware-practice.md). RAG/long-context workloads are poorly served; chat/code (small prompt, long decode) is the fit. llama.cpp/Ollama has faster prefill than MLX but slower decode on Apple Silicon ( — https://github.com/pitambarmahato/hardnumbers-experiments/blob/HEAD/mlx-vs-llama-cpp/README.md).
• Concurrency: naive servers (Ollama/llama.cpp server) effectively serialize — aggregate throughput plateaus (~200 tok/s regardless of users). With continuous batching (Velox, rapid-mlx), aggregate throughput scales: Velox measured 92.7 → 223 → 395 → 547 tok/s aggregate at 1/4/8/16 concurrent users (5.9× at 16 users); rapid-mlx at B=4 sustained concurrent got 176 tok/s aggregate on Qwen3.6-35B-A3B 4-bit vs 129 for stock mlx-lm serve ( — https://github.com/soflutionltd/velox ; https://github.com/hc-gbs/rapid-mlx). Economic planning assumption: 2.5–4× single-stream aggregate throughput with a proper batching server (conservative vs the 5.9× lab figure). Per-user latency degrades as concurrency rises; p95 stays flat-ish under batching vs collapsing linearly without it.
• Sustained-load thermals are fine: 70B-class inference holds full speed indefinitely on the Studio chassis (desktop thermals, unlike thin laptops where prefill throttles) ( — tbreak review; mlx-lab thermal note).

4. Tunneling / Networking

How Cloudflare Tunnel works for this scenario: the host runs the cloudflared daemon, which opens outbound encrypted connections to Cloudflare’s edge; a public hostname (custom domain on Cloudflare DNS, or a *.trycloudflare.com quick-tunnel URL) routes buyer API traffic back down the tunnel to the local inference server (e.g., mlx-lm serve on localhost:8000). No inbound ports opened, no port forwarding, NAT/CG-NAT traversed automatically (outbound-only). TLS terminates at Cloudflare’s edge (free cert); the tunnel leg is encrypted. Authentication is layered on top: Cloudflare Access policies, or — more appropriate for an API marketplace — API keys validated at the central gateway and/or per-host bearer tokens.

• Mechanics/cost: https://tech-insider.org/cloudflare-tunnel-setup-2026-homelab/ ; https://medium.com/@paperkite_hq/cloudflare-tunnels-zero-trust-access-to-your-self-hosted-services-7ce3dc67a5a5 ; https://github.com/joanmarcriera/home-office-automations/blob/HEAD/docs/services/cloudflare-mesh.md (read 2026-09-30)
• Cost: the tunnel connector itself is free with unmetered bandwidth; Zero Trust free plan covers up to 50 users; paid tiers ~$7/user/mo beyond that. cloudflared client is Apache 2.0 open source; the edge network is proprietary.
• ToS / commercial-resale issue — the key finding: Cloudflare’s subscription terms prohibit customers from selling, reselling, renting, or making the Service available to third parties for use on internet properties not owned by the customer ( — https://cfl.re/esa-2022-10-18). There is a separate partner/VAR reseller program for legitimate resale. Practical reading: each host running cloudflared under their own Cloudflare account to expose their own machine is arguably the host’s own use, not resale of Cloudflare’s service — the marketplace sells inference, not tunnel capacity. But (a) free-tier ToS was written for personal/small-team use and historically restricted heavy non-HTML traffic (restrictions “relaxed significantly” but not contractually guaranteed — quick tunnels especially), and (b) a marketplace routing all buyer traffic through Cloudflare free tier for a commercial API business is exactly the kind of use Cloudflare can throttle or terminate. Recommendation: do not build the business’s core transport on Cloudflare’s free tier. Either negotiate a paid/partner arrangement or own the transport (below).

Alternatives:

Recommended architecture: marketplace-owned WireGuard mesh or per-host tunnels to marketplace VPS gateways (transport you control, TLS + API-key auth + metering at the gateway), with Cloudflare Tunnel as a fallback onboarding path for hosts behind restrictive NAT. The gateway also solves routing, failover when a host goes offline mid-night, and per-token metering.

5. Economics

Cost side (per machine-night, M5 Max 128GB/1TB @ $5,399)

• Power: 7–9W idle; GPT-OSS 120B inference 101–115W sustained, 174W peak; Apple max continuous 480W ( — https://tbreak.com/mac-studio-m5-max-review-local-ai/ ; https://intelligibberish.com/articles/local-ai-power-and-thermals/, read 2026-09-30)
• Electricity rate 18.34¢/kWh: https://intelligibberish.com/articles/local-ai-power-and-thermals/ (EIA June 2026 via that source)
• Hardware amortization is 97% of nightly cost. Electricity is a rounding error. (Corroborating context: a “Sunk Cost” calculator found a $3,499 Mac Studio needs 44 years to break even when substituting hosted API spend — hardware dominates, electricity negligible: https://runtimewire.com/article/sunk-cost-local-llm-mac-studio-break-even)

Revenue side: 2026 market price per 1M tokens (comparable open models)

• — https://github.com/theopenco/llmgateway/commit/1267015607e1cdb9979cfde99d85be0f5b51604e (Sep 2026, OpenRouter-cross-checked) ; https://github.com/feather-store/feather/blob/HEAD/docs/research/inference-backend-recommendation.md (Q1 2026) ; https://github.com/ericrisco/rsc-harness/blob/HEAD/skills/together-fireworks/SKILL.md (read 2026-09-30)
• Fireworks published size-tier fallback (useful for unlisted models): <4B $0.10, 4–16B $0.20, >16B dense $0.90, MoE ≤56B $0.50, MoE 56.1–176B $1.20 per 1M tokens ( — via rsc-harness skill, docs.fireworks.ai)

Unit economics: tokens/night vs cost/night

8 hours = 28,800 s. Revenue scenarios at commodity market pricing, single-user vs batched (3× aggregate via continuous batching), 100% utilization:

• Break-even: cost $5.09/night requires, in the best realistic case (30B MoE, batched, 100% utilization, $2.24/night), 227% utilization — impossible. At a realistic 40–60% utilization the machine grosses ~$0.90–1.60/night vs $5.09 cost.
• Price needed to break even at best-case throughput: $5.09 ÷ 11.2M tokens = ~$0.45/1M blended, i.e. ~2.3× the commodity 30B-class price — only achievable with a genuine premium (privacy, dedication, SLA).
• Margin split (if it worked): typical marketplace take rates (Vast.ai-style) are 15–30%. On a $2.24/night gross at 20% take, the host keeps ~$1.79 — still 64% below the host’s amortized cost. The split is moot until gross clears cost.
• Sensitivity: 24/7 operation (not overnight-only) roughly triples revenue to ~$6.7–7.8/day vs ~$5.41/day all-in cost — approximately break-even at 100% utilization, but that is a datacenter-style operation, not the overnight-rental scenario, and leaves no return on capital or ops labor.
• The case that works — sunk hardware: for an already-owned Mac, marginal cost ≈ $0.16/night electricity. Gross $0.75–2.61/night → profit $0.60–2.45/night ≈ $18–74/month per machine at 100% utilization (halve it at 50%). Real pocket money for an idle machine; not a basis for buying hardware.
• For reference, decentralized GPU markets price an RTX 4090 (24GB) at $0.29/hr on Vast.ai ($2.32/8h) ( — https://aistoryland.com/best-ai-gpu-rental-platforms-for-businesses/), and hyperscalers charge 3–3.6× the cheapest spot GPU clouds for the same silicon ( — https://dev.to/fastgpu/what-it-costs-to-rent-an-h100-b200-or-rtx-4090-in-september-2026-live-prices-from-28-gpu-clouds-12n3) — but those rents are for raw compute by the hour, a different product than per-token inference.

6. Business Shapes & Risks

(a) Centralized marketplace (Vast.ai-style for Macs)

Hosts install an agent (model runner + batching server + WireGuard/tunnel client + metering); buyers get one API key at the gateway, which routes to hosts by capability tier (memory, model, measured tok/s), meters tokens, and splits revenue. This is the most natural shape and the only one that aggregates enough supply to be a real product. Requires: host attestation/benchmarking on onboarding, per-model capability tiers, failover when hosts sleep/wake, and a batching-capable server (stock Ollama serializes — §3).

(b) Bring-your-own-model private inference

User uploads private weights/LoRA; marketplace schedules them onto a host overnight; premium flat or per-token pricing. Higher willingness to pay (privacy-motivated), but directly collides with the privacy risk below: “private” inference on an untrusted stranger’s Mac is not private without confidential-computing guarantees Apple Silicon does not offer to third-party hosts.

(c) Other variants

• Dedicated flat-rate nightly node (“rent the box, 10pm–6am”): simpler metering, Vast.ai-like $/hr; buyer brings their own stack. Anchored by Vast.ai RTX 4090 ~$0.29/hr — a 128GB Mac Studio would need ~$0.60+/hr to clear its amortization, a tough sell against GPUs.
• Batch/off-peak arbitrage: embeddings, evals, distillation-data generation, benchmark sweeps — latency-insensitive work that fits the overnight window and tolerates host churn.
• Enterprise Mac-fleet monetization: companies with idle Mac fleets (CI farms, design studios) sell nights; solves trust partially (known operator) but they could just buy GPUs.
• Edge/cached inference: serve popular prompts at the edge — undercut on bandwidth, not tokens.

Key risks

1. Prompt data privacy on untrusted hosts — the #1 structural risk. The host’s OS sees plaintext prompts and completions. There is no TEE/confidential-compute story for third-party inference on Apple Silicon. Mitigations (reputation staking, contracts, prompt encryption) don’t fix the fundamental issue; confidential-computing or the host running inside the buyer’s trust domain is required for sensitive data. This caps the addressable market to non-sensitive workloads — exactly the workloads where buyers won’t pay a privacy premium.
2. Uptime/SLA. Consumer machines on residential ISPs, asleep/waking, owners reclaiming the machine. Overnight-only supply inverts normal demand (peak API demand is daytime US). Expect best-effort SLAs; enterprise buyers won’t accept them.
3. Open-weight license compliance for commercial hosting. Apache 2.0/MIT models (Qwen ≤32B tiers, GPT-OSS, DeepSeek, GLM) are clean for commercial hosting. Llama 4 Community License permits commercial use but requires “Built with Llama” attribution, Llama-prefixed naming for derivatives/fine-tunes, adherence to an Acceptable Use Policy, a 700M-MAU cap (won’t bind a startup, but acquirers’ lawyers will flag it), and excludes EU-domiciled companies from multimodal use ( — https://aifoss.dev/blog/open-weight-licences-that-block-commercial-use-2026/ ; https://automatenexus.com/blog/open-weight-model-licenses-explained ; license text https://github.com/meta-llama/llama-models). Kimi K3’s custom license needs redistribution review before shipping ( — medium.com benchmark article). Alibaba’s largest Qwen tiers have historically carried a separate Qwen License with a 100M-MAU commercial threshold — verify the exact checkpoint’s license file, not the family reputation ( — aifoss.dev).
4. Quality variance across heterogeneous hosts. Different chips (M1→M5), memory sizes, networks, and thermal environments → different models fit, different tok/s, different TTFT. Requires capability tiers and continuous re-benchmarking; buyers of “the API” will feel the variance as flaky latency.
5. Does hyperscaler/GPU pricing kill the arbitrage? Not directly — the killer is commodity inference providers, not raw GPU rents. DeepInfra serves Llama 3.3 70B at $0.23/$0.40 per 1M ( — feather research doc) — a price a $5,399 Mac physically cannot match per-token. The Mac’s only cost advantage is marginal cost on sunk hardware ($0.16/night), which no datacenter can touch — but that advantage accrues to the machine’s owner, not to a venture buying machines.

Bottom-line verdict

Is the arbitrage real? As a buy-hardware-to-sell-tokens business: no. The nightly math at 2026 commodity prices:

• Best realistic gross: ~$2.20–2.60/night (MoE model, batching server, 100% utilization, 8h)
• All-in cost: ~$5.09/night ($4.93 amortization + $0.16 electricity)
• Break-even needs >200% utilization or ~2.3× commodity token pricing

What nightly economics make it work? One of:

1. Sunk hardware (the actual scenario in the mission — people renting their own idle Macs): marginal cost ~$0.16/night; any utilization pays. Host nets ~$0.60–2.45/night at 100% utilization → $18–74/month/machine. Viable as a consumer side-hustle product, not as a hardware-investment business.
2. Premium pricing ≥$0.45/1M blended on the served tier (private BYOM, dedicated nodes, compliance-flavored hosting) — requires solving the trust/privacy contradiction first.
3. 24/7 operation at ≥80% utilization — approximately break-even on new hardware, but abandons the overnight-rental premise and still earns no return on capital.

Strategic note: the models that fit 128GB cap out at GPT-4o-class, so the product competes with DeepInfra/Together at $0.10–0.60/1M — a price war against datacenters that the Mac loses per-token and wins only on marginal cost. The defensible wedge is not cheaper tokens; it is monetizing already-idle consumer hardware (where marginal cost ≈ $0.16/night is unbeatable) plus software that makes heterogeneous Macs look like one reliable API (batching servers, capability tiers, failover, owned transport — not Cloudflare free tier).

What couldn’t be confirmed

• tbreak review’s exact benchmark/power figures: the live-browser read of https://tbreak.com/mac-studio-m5-max-review-local-ai/ failed with a terminal observation error on 2026-09-30; figures are corroborated by independent sources (mlx-lab, baem1n/llm-bench) but not directly re-verified.
• Concurrent-serving throughput on the M5 Max specifically: Velox/rapid-mlx concurrency figures were measured on other Apple Silicon machines (M3 Ultra etc.); the 2.5–4× batching multiplier applied to M5 Max is an engineering estimate, not a measured M5 Max result.
• Exact per-checkpoint license files for the recommended models (Qwen3.6-35B-A3B, GPT-OSS 120B, Mistral Small 4, Qwen3.5-122B-A10B) — verify each Hugging Face checkpoint’s LICENSE file before commercial hosting; family-level reputation is not sufficient (Qwen’s large tiers have historically differed).
• Cloudflare’s current free-tier enforcement posture toward commercial API traffic — ToS text reviewed ; no live test or Cloudflare statement obtained. Treat as a business risk, not a settled permission.
• Real buyer demand / willingness to pay for Mac-hosted inference at any price — no marketplace of this kind was found operating at scale; demand is assumed, not evidenced.

Sources

All sources read 2026-09-30.

Hardware & pricing

• https://www.apple.com/shop/buy-mac/mac-studio (configurator: $2,499 base; +$300 40-core GPU; +$2,000 128GB; $48.99/mo lease)
• https://pub.towardsai.net/m5-max-128gb-or-m5-ultra-96gb-the-mac-studio-buying-guide-for-local-llms-in-2026-5b7171d482af — ($5,399 for 128GB/1TB)
• https://wavect.io/blog/mac-mini-m6-vs-mac-studio-m5-local-ai/ — (specs, bandwidth, tiers)
• https://www.macobserver.com/news/mac-studio-m5-max-vs-m5-ultra-configurations-compared/
• https://www.macworld.com/article/2973459/2026-mac-studio-m5-release-date-specs-price-rumors.html
• https://forums.appleinsider.com/discussion/245832 Throughput & power
• https://tbreak.com/mac-studio-m5-max-review-local-ai/ — (Qwen3 30B-A3B 130.6 tok/s gen; Llama 3.3 70B 13.1 tok/s; GPT-OSS 120B 75.6 tok/s sustained; 101–115W inference power)
• https://github.com/suryals/mlx-lab — (70B 11 tok/s / 39.9GB; 123B 6.3 tok/s / 69.2GB)
• https://github.com/baem1n/llm-bench (via https://github.com/1solon/home-server-configuration/blob/HEAD/docs/research/m5-ultra-local-llm-performance.md) — (Qwen3.5-35B-A3B 139 tok/s)
• https://www.youtube.com/watch?v=MwBSNUur1B0 — (Qwen3 Coder Next 80B 75 tok/s)
• https://github.com/soflutionltd/velox — (continuous batching: 5.9× aggregate at 16 users)
• https://github.com/hc-gbs/rapid-mlx — (B=4 concurrent: 176 tok/s aggregate Qwen3.6-35B-A3B)
• https://github.com/t0rsion/leone/blob/HEAD/research/07-consumer-hardware-practice.md — (prefill weakness; concurrency notes)
• https://github.com/pitambarmahato/hardnumbers-experiments/blob/HEAD/mlx-vs-llama-cpp/README.md — (MLX vs llama.cpp prefill/decode tradeoffs)
• https://intelligibberish.com/articles/local-ai-power-and-thermals/ — (480W max continuous; 18.34¢/kWh US avg residential Jun 2026)
• https://runtimewire.com/article/sunk-cost-local-llm-mac-studio-break-even — (44-year break-even context)

Models

• https://codersera.com/blog/best-open-source-llm-2026-llama-4-qwen-3-5-deepseek-v4-gemma-4-mistral/
• https://medium.com/@surajkhaitan16/i-benchmarked-the-top-20-llms-of-2026-heres-which-to-use-for-what-f2447efc5954
• https://tech-insider.org/llama-4-vs-qwen-vs-mistral-2026/
• https://callsphere.ai/blog/llm-comparison-code-generation-open-vs-open-may-2026
• https://github.com/wty2003328/silicon-to-serving/blob/HEAD/L6_Algorithms_and_Models/07_Frontier_Models_2025_2026.md
• https://github.com/neo-x/llm-playground/blob/HEAD/LLMs%20performance%20analsys.md — (70–80B class ≈ GPT-4o)

Pricing (inference market & GPU rental)

• https://github.com/theopenco/llmgateway/commit/1267015607e1cdb9979cfde99d85be0f5b51604e — (DeepSeek V4.1 Flash: DeepInfra $0.20/$0.60, Fireworks $0.22/$0.66, Together $0.30/$1.20)
• https://github.com/feather-store/feather/blob/HEAD/docs/research/inference-backend-recommendation.md — (Llama 3.3 70B: Together $0.88, Fireworks $0.90, Groq $0.59/$0.79, DeepInfra $0.23/$0.40)
• https://github.com/ericrisco/rsc-harness/blob/HEAD/skills/together-fireworks/SKILL.md — (Fireworks size-tier fallback pricing)
• https://aistoryland.com/best-ai-gpu-rental-platforms-for-businesses/ — (Vast.ai RTX 4090 $0.29/hr)
• https://dev.to/fastgpu/what-it-costs-to-rent-an-h100-b200-or-rtx-4090-in-september-2026-live-prices-from-28-gpu-clouds-12n3 — (hyperscaler 3–3.6× spot gap)

Tunneling & ToS

• https://tech-insider.org/cloudflare-tunnel-setup-2026-homelab/ — (free tunnel, unmetered bandwidth, 50-user free plan)
• https://medium.com/@paperkite_hq/cloudflare-tunnels-zero-trust-access-to-your-self-hosted-services-7ce3dc67a5a5
• https://github.com/joanmarcriera/home-office-automations/blob/HEAD/docs/services/cloudflare-mesh.md
• https://github.com/noahwins-ng/equity-data-agent/blob/HEAD/docs/decisions/018-cloudflare-quick-tunnel-for-https-ingress.md
• https://cfl.re/esa-2022-10-18 — (Cloudflare terms: no resale/rental of the Service to third parties)
• https://instatunnel.my/blog/navigating-the-2026-tunneling-landscape-budget-performance-and-the-end-of-the-ngrok-monopoly — (ngrok 2026 pricing: free 1GB/mo; paid $8+/mo, 5GB then $0.10/GB)
• https://Slashdot.org/software/comparison/Cloudflare-Tunnel-vs-Tailscale/ — (Tailscale ~$5/user/mo paid)

Licenses

• https://aifoss.dev/blog/open-weight-licences-that-block-commercial-use-2026/ — (Llama 4: 700M MAU cap, attribution, EU multimodal exclusion; Qwen large-tier license caveat)
• https://automatenexus.com/blog/open-weight-model-licenses-explained — (Apache 2.0/MIT = clean; Llama community license strings)
• https://github.com/onehillai/aoi/blob/HEAD/models/meta-llama/assess.md — (Llama license assessment)