A 700-watt chip beat a 1,400-watt one. That's the number a product manager sees first, and it's the number that gets Jalapeño into the meeting.
OpenAI benchmarked its first custom inference chip against Nvidia's Blackwell systems on SemiAnalysis's InferenceX suite and, per OpenAI's own figures, posted 1.5x to 1.9x more work per watt, 1.7x to 3.6x lower end-to-end latency, and 2.1x to 4.1x higher performance on highly interactive workloads that matter most to anyone building today. Reportedly at roughly half the power draw. On a first-generation chip, from a company that had never taped out silicon before.
The efficiency win is real. SemiAnalysis went to OpenAI's labs, ran the benchmark in person, and confirmed it beats Blackwell on perf-per-watt in nearly every scenario tested — and, per Jon Peddie Research, at parity on total cost of ownership against Nvidia's newer Rubin, without speculative decoding turned on. This is not vaporware.
So here's the only question that matters if you have an OpenAI line item: does captive silicon show up in my pricing, or just in their margin?
What developers actually get right now: nothing. Not directly.
Broad deployment doesn't arrive until 2027. And OpenAI has made no commitment to pass any of the savings through to API pricing.
Read the savings language closely, because it's all pointed inward. Richard Ho, OpenAI's head of hardware, told Bloomberg the 700-watt draw "should reduce data center power costs." His exact words on the savings: "It's a really good chip — it should drop it by a lot." Broadcom reportedly cited early samples showing roughly 50% cost savings versus typical GPUs.
Notice what none of those sentences mention: your invoice. Every savings claim describes OpenAI's cost structure. Not one describes developer pricing. That gap is the whole story.
Now the reason the gap is likely to stay a gap. SemiAnalysis and Jon Peddie both flag the same constraint: OpenAI designs for perf-per-watt because it is currently limited by datacenter power, not by budget. Grid interconnection moves slower than hardware deployment. The scarce resource is the watt, not the dollar.
That single fact predicts where the efficiency gain goes. If your bottleneck is power and you suddenly get 1.9x more work per watt, you don't lower prices — you serve more customers with the power you already have. The efficiency gets eaten by OpenAI's own capacity hunger before it ever reaches a price cut. Internal demand can't even be routed outward today; that's the definition of a supply-constrained business absorbing its own gains.
If your bottleneck is power and you suddenly get 1.9x more work per watt, you don't lower prices — you serve more customers with the power you already have.
So let's rank the four things OpenAI floated, from most to least likely.
Most likely: higher rate limits at existing price tiers. This is the cleanest fit for a power-constrained operator. More capacity per megawatt directly translates to more requests served, and OpenAI can grant that without touching the price sheet or admitting anything about its cost base.
Second: faster response times on certain models as Jalapeño capacity grows. The chip's whole architecture — local KV cache, no prefill-decode disaggregation, reportedly over 700 tokens per second per user at concurrency 1 on DeepSeek R1 — is built for latency. Passing through speed costs OpenAI nothing in negotiating leverage. It's a feature, not a discount.
Third: stable API pricing while competitors raise rates. This is subtle. It's not a price cut; it's the absence of a price increase, framed as a benefit. But "we didn't raise prices" is a defensive moat, not a gift to you.
Least likely: lower-cost long-context and agent workloads showing up as cheaper API pricing. This is the one PMs want, and it's the one that requires OpenAI to voluntarily hand back margin on a chip nobody else can run, in exactly the workload category — agents, million-token context — that OpenAI has said it expects to grow fastest. You do not discount your fastest-growing, most-differentiated product while power-constrained. You harvest it.
There is a mechanism by which a real pass-through could appear, and it's worth naming. Per QZ, Ho said the chip's ability to excel at both volume-heavy and speed-sensitive workloads means OpenAI can offer customers a choice between deployments tuned for lower cost or for faster answers. That's the tiered-inference lever — the same time-of-day and deployment-tier pricing thread that's been open for a while now. Jalapeño gives OpenAI the cost headroom to build those tiers. Whether it uses that headroom to lower your floor or just to widen its own menu is, again, entirely OpenAI's call.
Here's the part that should actually change how you think, beyond the pricing timeline.
Captive silicon widens the gap between two paths. On one side, you rent OpenAI's chip — except you can't rent it, you rent the API in front of it. There is no second provider running Jalapeño. There is no spot market, no floor to negotiate against, no merchant-silicon comparable to price-check. OpenAI, together with Broadcom, effectively controls the chip, the model, the serving stack, and the price sheet end to end.
On the other side sits the DeepSeek and open-model path. Jalapeño's own benchmarks ran on DeepSeek R1 and Kimi K2.5 — models you can take to any hardware vendor. On those paths you can at least shop the silicon. You can put the same weights on Nvidia, on AMD, on a neocloud, and force a bid.
That's the real lock-in mechanism, and it's dressed up as a performance story. A lab owning its own chip means its cost structure becomes a black box. You can no longer back into OpenAI's economics by pricing merchant GPUs, because OpenAI isn't paying merchant-GPU prices anymore. When your vendor's costs are unbenchmarkable, your leverage in every future pricing conversation quietly shrinks.
So the decision rule.
Do not reprice your product or renegotiate your OpenAI contracts on Jalapeño today. The chip is an engineering sample, deployment is small-volume through late 2026, and nothing about your invoice has changed. The honest timeline for any developer-facing effect is 2027 at the earliest.
Watch one signal, and only one: rate-limit and price announcements in Q1 2027. That's the first window where deployed Jalapeño capacity could plausibly touch developers. If what arrives is higher rate limits — expect that. If what arrives is an actual per-token price cut on agent and long-context workloads — that's the surprise, and it would tell you OpenAI chose to compete on price rather than harvest margin.
Bet on the former.
The benchmark is genuine, and the inflection is genuine — a frontier lab that beats Nvidia on efficiency stops being purely a customer and starts being a competitor for its own workloads. But that shift changes who captures the margin in AI infrastructure. It says nothing about that margin landing in your favor. The most likely outcome of the four things OpenAI dangled is the one it never quite promised: the savings stay inside the building.