RTX 3090 still wins on privacy and marginal cost.
| Cloud API today |
Wait for Jalapeño savings |
Used RTX 3090 local |
| Cost (10M tokens/mo) |
~90ドル–100ドル/mo |
Unknown, not before 2027 |
~50ドル/mo all-in (yr 1) |
| When available |
Now |
Prototype end-2026, ramp 2027–28 |
Now (~1,070ドル) |
| The catch |
Per-token price set by OpenAI, not its silicon cost |
API price ≠ wafer cost; no committed pass-through |
1,070ドル up front, you run the rig |
Honest take: Jalapeño is a Wall Street story, not a home-lab one. If you were going to buy a used RTX 3090 this month, buy it — nothing announced on June 24 makes waiting the smarter move.
What OpenAI and Broadcom actually announced
On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, described as OpenAI's first "Intelligence Processor" — a custom ASIC built specifically for large language model inference rather than the general-purpose work a GPU has to handle. CNBC, TechCrunch, Tom's Hardware, and VentureBeat all covered the launch the same day, alongside OpenAI's and Broadcom's own press releases.
The technical claims worth pinning down:
-
It's an inference chip, not a training chip. Jalapeño is tuned around the memory movement, kernel, and networking patterns that matter for transformer inference. It is not meant to replace the GPU fleet OpenAI uses for training and experimentation.
-
~50% lower cost per token vs current NVIDIA GPUs. This is the headline number (reported by TechTimes), and it's self-reported — based on OpenAI's own workloads, without a disclosed comparison baseline or independent verification.
-
Performance per watt "substantially better than current state-of-the-art." OpenAI says a detailed technical report is coming in the following months. As of launch day, there were no published tokens/sec or watts figures.
-
Built in nine months, manufactured by TSMC. OpenAI calls the design-to-tape-out cycle possibly the fastest ever for a high-performance ASIC, and credits its own models with speeding up parts of the design work. The die is reticle-sized.
-
Deployment timeline: prototype by end of 2026, production ramp in 2027 and 2028. Gigawatt-scale data centers are planned with Microsoft and other partners. The reported deal with Broadcom is valued around 10ドル billion.
That's the whole picture: a memory-bottleneck-focused inference ASIC that lowers OpenAI's own cost to serve frontier models, deploying slowly over the next two years.
The leap the headlines want you to make (and shouldn't)
The implied argument in a lot of the coverage is: cloud inference is about to get much cheaper, so why buy hardware? Three things break that chain before it reaches your wallet.
1. Cheaper for OpenAI ≠ cheaper for you. A 50% cut in cost-per-token is a statement about OpenAI's cost to serve, not about the price on the API menu. Those are different numbers set by different forces. OpenAI raised GPT-5.5 to 5ドル.00 per million input tokens and 30ドル.00 per million output tokens when it launched on April 23, 2026 — double the previous GPT-5 line — at a time when its NVIDIA-based serving costs were presumably already falling with Blackwell. Pricing tracks competition, demand, and margin targets, not the bill of materials. A company that just doubled token prices is not the company that reflexively passes silicon savings to developers.
2. The timeline is 2027–2028, not now. Jalapeño hits "prototype deployments" by the end of 2026 and ramps through 2027 and 2028. Even in the optimistic case where some of the savings reach API prices, that's a 2027-at-the-earliest event, and only after the chip is serving meaningful volume. You'd be deferring a decision you can act on today against a discount that may never be itemized.
3. It doesn't touch privacy or offline use. The entire reason a large slice of this site's readers run models locally is that the data never leaves the machine. No inference ASIC in an OpenAI data center changes that. If your use case is "summarize my private notes" or "code against a proprietary repo without sending it anywhere," the cloud price could go to zero and local would still win.
The actual cost math, run honestly
Here's the comparison that matters for an indie dev or home-labber doing real volume — say 10 million tokens a month, a realistic figure for daily coding assistance, document Q&A, and drafting.
Cloud (GPT-5.5, ~80% input / 20% output split):
- 8M input ×ばつ 5ドル.00/M = 40ドル.00
- 2M output ×ばつ 30ドル.00/M = 60ドル.00
-
≈ 100ドル/month (less if you lean on the 90% cached-input discount of 0ドル.50/M for repeated context)
For reference, Claude Opus 4.8 runs 5ドル/25ドル per million and lands near 90ドル/month on the same split; Claude Fable 5 at 10ドル/50ドル roughly doubles that. None of these are Jalapeño-affected — they're NVIDIA-served today and priced on competition.
Local (used RTX 3090):
- Card: ~1,070ドル used in June 2026 (lowest monthly average; the broader market average across 338 listings sits around 1,237ドル). Amortize 1,070ドル over 24 months = 44ドル.58/month.
- Electricity: the 3090 draws about 350W under inference load and ~21W idle. Run it 4 hours a day at load = ~1.4 kWh/day, ~42 kWh/month, ~5ドル/month at 0ドル.12/kWh.
-
≈ 50ドル/month in year one, dropping to ~5ドル/month once the card is paid off.
So even before Jalapeño, local already wins on a two-year horizon for steady usage — roughly 50ドル/month vs ~90ドル–100ドル/month — and the gap widens every month after the card amortizes. A hypothetical future cloud discount has to overcome a head start, not a deficit. (If your usage is bursty or you only need a big model occasionally, the calculus flips toward renting — see RunPod vs local GPU for where the break-even actually lands, and our 400ドル/month GPU bill breakdown for how indie devs overspend on cloud.)
The used 3090 holds up here for the same reason it holds up everywhere on this site: 936 GB/s of memory bandwidth and ~95 tokens/sec on a 7B model at Q4, for the price of a year of moderate API use. Its full case is in Used RTX 3090 in 2026: Still the AI Value King?.
Where Jalapeño does matter (just not to you yet)
This isn't to dismiss the chip. It's a genuinely big deal for the industry:
-
It pressures NVIDIA's pricing power. OpenAI joining Google (TPU) and Amazon (Trainium) in building custom inference silicon chips away at NVIDIA's near-monopoly on AI compute margins. Over years, that could drag down the cost floor for everyone — including the cloud providers you rent from.
-
It validates the "inference is memory-bound" thesis. Jalapeño targets the memory bottleneck rather than piling on FLOPS, which is exactly the lesson home-labbers learned the hard way: a used 3090's bandwidth beats a newer card's TOPS for token generation. The same physics that makes Jalapeño efficient is why bandwidth, not compute, governs your local tokens/sec.
-
It's part of a broader custom-silicon wave. It rhymes with what's happening at the accessible end of the market, too — see Qualcomm's 10ドルB Tenstorrent bid, where RISC-V AI cards are actually buyable for home labs today, unlike Jalapeño, which you will never be able to put in a PCIe slot.
But none of that is a 2026 home-lab purchasing input. It's a multi-year m