Bochinski lays out a concrete TCO: ~$1,400 for dual used RTX 3090s, 128GB DDR5, and a 1600W PSU, plus ~$32/month in Pittsburgh power costs. Against a $200/month Claude Code subscription, the rig breaks even in seven months and continues paying out across the four-year life of the cards, even after accounting for UPS, storage, and SRE tax.
Argues the open-weight gap against Sonnet 4.7 and GPT-5.1 effectively closed in Q1 2026 for the workloads most devs actually do — refactors, test generation, code review, and small-scope feature work. Points to Qwen3-Coder and DeepSeek-V3.1-Coder landing within a couple of points of frontier models on SWE-Bench Verified, with GLM-4.6 as the consensus pick for agentic loops.
Bochinski explicitly calls out the parts of the bill hardware reviewers omit: an hour spent flashing BIOS for PCIe bifurcation, a UPS, a second SSD for model storage, and the recurring tax of llama.cpp shipping breaking changes. He frames self-hosting as a real operational commitment rather than a one-time hardware purchase.
The editorial concedes the raw model gap has closed for common workloads, but argues what hasn't closed is the system gap — the agentic loops, integrations, and polished harness that hosted Claude Code provides out of the box. Self-hosters get comparable model quality but must assemble their own scaffolding around Continue, Aider, and Zed.
Stephen Bochinski's post ‘AI Coding at Home Without Going Broke’ hit 277 on Hacker News on June 13 by doing the one thing this audience has been asking for: a no-bullshit TCO breakdown of running coding assistants on your own metal in 2026. Not a benchmark dump, not a hardware-porn build log — an actual answer to the question every senior dev has muttered after their Anthropic invoice clears: *is the local stack good enough yet, and what does it really cost?*
Bochinski's setup is unsexy on purpose: two used RTX 3090s on a consumer X670 board, 128GB DDR5, a 1600W PSU, total bill ~$1,400 after eBay scavenging. He runs Qwen3-Coder-30B-A3B at 6-bit quantization through llama.cpp with the new speculative-decoding pipeline, fronted by an OpenAI-compatible endpoint that drops into Continue, Aider, and Zed without modification. Idle draw is 90W; under load it sits at 480W and pulls roughly $32/month at his Pittsburgh power rate.
Against a $200/month Claude Code subscription, the rig breaks even in seven months and keeps paying out for the four-year life of the cards. He's careful to add the parts of the bill that hardware reviewers always forget: a UPS, a second SSD for model storage, the hour he spent flashing a BIOS to get PCIe bifurcation working, and the recurring tax of being your own SRE when llama.cpp ships a breaking change.
The HN thread is the real story. The top comment, from someone running a Mac Studio M3 Ultra, points out that the open-weight gap against Sonnet 4.7 and GPT-5.1 effectively closed in Q1 2026 for the workloads most devs actually do — refactors, test generation, code review, small-scope feature work. Qwen3-Coder and DeepSeek-V3.1-Coder both land within a couple of points of the frontier on SWE-Bench Verified, and GLM-4.6 is now the consensus pick for agentic loops where you need long-horizon tool calling without hallucinated function names.
What hasn't closed is the *system* gap. Hosted Claude Code isn't just a model — it's a model plus a battle-tested agent harness, plus tool routing, plus a sandboxed shell, plus context compaction that actually works at 200K+ tokens. Running Qwen locally gets you the brain; you're still on the hook for the nervous system. Aider is the closest thing to parity for autonomous edits, but every commenter who's tried to wire local models into a real agentic workflow ends up writing their own scaffolding — and discovers, painfully, why Anthropic charges what it does.
The second thread of disagreement is about hardware tiers. The 2× 3090 build is the obvious sweet spot, but the prices are creeping up as gamers and hobbyists chase the same cards. A user named ‘mlserv’ argues the Mac Studio M3 Ultra with 256GB unified memory is the better long-term bet — slower tokens-per-second, but it runs the 235B Qwen3-Coder-Max model that the dual-3090s can't fit. The counter from ‘nh2’: at $7,000, the Mac breaks even in fourteen months against the same Claude subscription, and you've still got no upgrade path because Apple solders the RAM.
Then there's the dark horse: the AMD Strix Halo mini-PCs that shipped in volume in March. 128GB of unified LPDDR5X, an RDNA 3.5 iGPU with the new ROCm coder-tuned kernels, and a $1,900 price tag. Tokens-per-second is half of a 3090, but the power draw is 65W under load — which, over four years in California, saves more than the hardware delta. Two HN commenters report they're running 70B-class models on Strix Halo with no fan noise and no second PSU. That's the form factor that makes self-hosted coding assistants feasible for people who don't have a basement.
The deepest comment in the thread is from a former Cursor engineer who points out that the economics flip again if you account for *team* usage. A single 3090 rig serving four developers via tailscale costs the same as one developer's subscription but replaces four — and the marginal hardware cost of adding a fifth dev is zero until you saturate the GPU. This is the angle nobody's pricing correctly: the hosted players charge per-seat, but inference is fungible.
If you write code for a living and you've been paying $200+/month to Anthropic or OpenAI for eighteen months, the math has flipped. Buy used 3090s while you still can, install llama.cpp's server, point Aider or Continue at it, and budget a weekend for setup. You will not get parity with Claude Code on long-running agent tasks — keep the subscription for those — but you will get 80% of your daily coding assist for $32/month in electricity.
If you're a two-to-six-person team, the rig pays for itself in under three months. Tailscale, a wildcard cert, and an Nginx that does simple round-robin between two endpoints is a Saturday's work. The hidden upside: your code stops crossing somebody else's training set, which is a non-trivial answer when your customers ask about it.
The mistake to avoid is treating this as a binary swap. Keep the hosted subscription for agentic, multi-file, multi-hour tasks where the harness matters more than the model. Use local for the inner loop — autocomplete, single-file refactors, test scaffolding, code review on diffs. Bochinski's own punchline is that his Anthropic bill dropped 78%, not 100%, and he considers that the right outcome.
The next twelve months are about the harness, not the model. Whoever ships an open-source agent runtime that matches Claude Code's tool routing and context management will collapse the remaining gap, and the frontier labs will be forced to compete on something other than the assumption that nobody can run their weights. The hardware curve helps: Strix Halo refreshes are due in Q4, NVIDIA's rumored 5070-Super-with-32GB exists for exactly this market, and used H100s have started leaking out of failed AI startups at $12K a pop. Self-hosted coding stopped being a hobbyist position this year. By next summer it'll be the default for anyone paying their own bill.
> The first is to self host. You buy the machine, run open source models locally, and pay nothing per token after that.Power is not free.What I’ve found is that you’re basically paying a premium for privacy, and that’s worth it for me.
I cannot figure out what people are doing to spend all this money.I have used a $60 per month Cursor plan on auto, and have never come close to using up my included usage, and I probably have it planning and coding and working for me all through the evenings 4 nights a week.What on earth are people
> The upfront cost is steep and the models you can actually run at home are weaker than what the frontier labs ship, so this only pays off if you can keep the rig busy with long running tasks where a slower, cheaper model grinds away overnight. Most people can’t keep a home machine that loaded, a
I find just going via Deepseek's platform API directly, using their V4 flash model, and hooking into a harness like Opencode more than acceptable. Think I've spent maybe $10 over a couple of weeks.I did explore self-hosting models but hardware right now is just too expensive.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I feel like I must have plateued and don't know what to do next to level up. I'm currently on the $100/month codex plan and it seems fine using 5.5-xhigh all the time. I think of what to do next, have a chat session to determine exactly what to ask for up to the point of being ready t