AI Designs Chips: Five Projects, but Only One Has Measured Silicon
Within a few days, OpenAI, the startup Architect Labs and a solo developer all announced records with AI-designed chips: 14 percent, 4.59x, 33x. We ranked the cases by maturity, from simulation to finished chip. The pattern: the big numbers usually come from low on that ladder, and every multiplier hides a specific comparison.

What this is about
In the space of three days, several announcements claimed AI is designing computer chips. On November 18, OpenAI said its own models had improved circuits in AMD's flagship GPU by “up to 14 percent.” The same day, startup Architect Labs unveiled Frostbite, a CPU designed by AI agents that is supposedly 4.59 times more efficient per watt. On November 20, developer Ankit Gupta described on his blog how coding agents sped up an LLM accelerator on an Amazon FPGA 33-fold, despite his having no hardware experience himself. Add to that the open GitHub project apex-inference-chip and the Design Conductor 2.0 paper from startup Verkor, published back in May.
The numbers sound impressive, but they aren't comparable. On its way to a finished product, a chip passes through several stages. The lower a project sits, the more can still change. So we've placed the five cases on a maturity ladder and checked what each multiplier actually refers to.
The ladder: from simulation to measured chip
The stages, briefly explained:
- RTL simulation: The circuit is written as code (RTL, Register Transfer Level, a kind of blueprint in languages like Verilog) and simulated on a computer.
- FPGA: The circuit runs on a reprogrammable chip, i.e. on real hardware. But clock speeds and efficiency are far below those of a custom-fabricated chip.
- GDS: The finished layout exists, the file a chip fab receives for manufacturing. Power and clock figures are still calculated, though.
- Tape-out: The design has been sent off for fabrication.
- Measured silicon: The finished chip runs and has been measured.
| Project | Highest stage reached | Headline number | Baseline |
|---|---|---|---|
| OpenAI × Jalapeño | 5 – silicon (function confirmed) | approx. 13 % / 10 % less cell area | own baseline designs, values estimated from synthesis |
| OpenAI × AMD GPU | 4 – tape-out (individual changes) | up to 14 % PPA | highly optimized AMD blocks |
| Frostbite (Architect Labs) | 3 – GDS (7 nm, no FPGA) | 4.59× performance per watt | an unnamed in-order core |
| Gupta, FPGA accelerator | 2 – FPGA, full model | 33× | own first version (6.43 tok/s) |
| Design Conductor 2.0 (VerTQ) | 2 – FPGA, 125 MHz | 80 hours of development time | – (16 nm area only estimated) |
| apex-inference-chip | 2 – FPGA, partial functions | 0.56 tok/s | own initial version (0.004 tok/s) |
33x – but compared to what?
Gupta has published the cleanest data, which is exactly why it's worth running the numbers. On an AWS F2 instance ($1.98 per hour, 16 GB of HBM), a Qwen-like model ran with 24 layers and ternary weights (only the values −1, 0 and 1). The first complete version that was bit-exact, meaning it produced exactly the same outputs as a CPU reference, managed 6.43 tokens per second. By the end it was 213.13. That's the 33x. The accelerator was built in 14 measured steps, each delivering roughly 1.5x according to Gupta. In total it took 166 prompts, only seven of them for steering. Across the whole series of experiments, the agents (Claude Opus and OpenAI models) wrote around three million lines of code and analyzed 30 million lines of output files. The tokens cost about $10,000.
The real benchmark is in Gupta's own table. A CPU server with 32 Zen 5 threads (m8a.8xlarge, $1.95 per hour) reaches 129.5 tokens per second with an exact CPU replica. Against this equally priced machine, the FPGA is only 1.65 times faster. Our calculation with his numbers goes further: the small 4-core m8a.xlarge instance manages 67.6 tokens per second for $0.24 per hour. That's around 282 tokens per second per dollar of hourly cost, versus around 108 for the FPGA. Per dollar, the small CPU comes out a good 2.6 times ahead. Gupta himself writes that a like-for-like comparison is difficult. He didn't point the agent loop at the alternatives, where there's probably more to gain. The write-up doesn't say who built the CPU version or how thoroughly it was optimized.
The setbacks are instructive too. Two hangs only showed up in hardware. A clock of 231.25/450 MHz produced random errors in a soak test (the system runs under load for hours), so the final version runs at 225/425 MHz. One major integration cost 38 hours because bugs a simulation could have caught only surfaced on the board the next day. With long contexts, speed drops from around 233 to about 76 tokens per second. Gupta's conclusion is sober:
“Humans are still useful for defining how to iterate quickly and what correctness looks like.”
He explicitly doesn't recommend the solution for production use. In the Hacker News discussion, FPGA practitioners point out that with FPGAs you pay for reprogrammability in cost, speed and power. The overhead compared with a custom-fabricated chip is 10 to 100x.
Frostbite: finished layout, secret opponent
Architect Labs sits higher on the ladder than Gupta. There's a finished GDS layout in 7 nm that, according to the company, hit 800 MHz after synthesis and 500 MHz in the final layout. A quad of four compute tiles plus four L3 nodes takes up 0.792 mm², the comparison core 0.895 mm². But Frostbite never ran on an FPGA, and the chip wasn't fabricated either. The power figures come from the company's own system model, Arcmodel. Before calibration it deviated from the circuit by up to almost 250 percent; afterward, according to the company, by 1 to 5 percent. The company itself concedes these are “of course still pre-silicon numbers.”
The 4.59x figure is an average (geometric mean) across 25 test programs. On the 10 programs the agents optimized for, it's 6.02. On the 15 held-out programs it's 3.83, and that's the more honest number. Depending on the program, the range runs from 23.12 (conv2d) to 1.15 (Fibonacci). The comparison is a single quad against a single core, a “modern superscalar dual-issue in-order core.” The company won't say which core “for legal reasons.” In-order means the core executes instructions strictly in sequence. That's a modest opponent, not laptop or server class. explainx.ai and a commenter on dev.to also point to tiny test tasks, such as mergesort over 1,024 elements or Fibonacci of 20. Architect Labs gives no date for a tape-out. With its predecessor project Redwood, the measured FPGA figure of 12.1 tokens per second was, according to VentureBeat, below Nvidia's Jetson Orin Nano (28). Only the projection to an unbuilt 8 nm chip produced a 3.4x per-watt lead.
OpenAI and AMD: highest on the ladder, fewest details
OpenAI is the only case with real silicon. Since early 2025, starting with GPT-o3 mini, its models have been helping design the company's own inference chip, Jalapeño. The changes were confirmed in silicon at the chip's first bring-up. The team estimates they reduced standard cell area (the area of prefabricated logic building blocks) by about 13 percent in the vector unit and about 10 percent in the matrix unit. A newly designed multiplier needed around 15 percent less area in synthesis and one fewer pipeline stage. So “proven in silicon” here means: the circuit works. The percentages are still estimates from the design process.
At AMD, the work involved individual blocks. OpenAI's exact words are “up to 14 percent improvement over highly optimized baseline designs on flagship GPU IP.” Whether that means power, performance or area remains unclear, as does which blocks were involved. According to OpenAI, some changes are in a design AMD has sent to fabrication. Nothing from it has been measured yet. For an upcoming CPU, AMD reportedly verified single-digit percentage gains in key blocks. The report gets more specific on a benchmark using the open RISC-V cores PicoRV32 and Ibex (9,881 runs). But there, the synthesis tool Yosys does the scoring, and timing verification after physical design is missing. OpenAI acknowledges this itself. Several commentators stress that all the numbers come from OpenAI itself and haven't been independently verified. OpenAI also has a business relationship with AMD.
Open projects: apex-inference-chip and Design Conductor 2.0
The repo SigmanticAI/apex-inference-chip (Apache-2.0) is remarkably transparent about its evidence. It maintains machine-generated status lists and a traceability file for known gaps, and labels every number as measured or projected. A single transformer decoder layer with a compressed KV cache (the buffer for already-processed tokens) has been implemented in RTL. In simulation it computes bit-exactly against a NumPy reference model. On the AWS FPGA, the project measures 0.56 tokens per second with Qwen2.5-0.5B at 62.5 MHz. The part that controls attention on the FPGA itself is still being brought up. All claims about 7B models, such as read speed at around 3 watts, are projections. There's no silicon; a completion for the open Sky130 fabrication process is on the roadmap.
Verkor's Design Conductor 2.0 (arXiv 2605.05170) produced the VerTQ accelerator for Google's TurboQuant compression method in 80 hours. It ran on a Xilinx FPGA at 125 MHz. Testing was done in system simulation with Qwen3-4B, but with only 64 tokens of context. The 5.7 mm² in TSMC's 16 nm process is explicitly an estimate. The paper shows a GDS only for an AES encryption core in the academic ASAP7 process. According to the authors, verification and timing closure, i.e. proving that all signals arrive in time, were the most expensive phases. The biggest bottleneck, they say, is human review.
Where agents already save work today
The evidence shows a pattern. Agents are strong at trying lots of variants when there's a hard automatic checker: bit-exactness for Gupta, equivalence checking with Synopsys tools for OpenAI, Lean proofs for Frostbite. That's exactly what the big EDA vendors (makers of chip design software) are selling. According to Tom's Hardware, Cadence cites 40x faster RTL validation at Nvidia, and Synopsys two to five times faster from spec to verified RTL. These are vendor or customer claims with differing baselines, though. None of the three big vendors' agents is demonstrably generally available yet.
Google's AlphaChip serves as a cautionary tale. The method demonstrably placed blocks in the fabricated TPU generations v5e, v5p and Trillium, with 3.2 and 6.2 percent shorter wiring respectively, according to Google. Yet experts have been arguing since 2021 over whether it beats established methods. According to Wikipedia, there's still no positive independent replication in peer-reviewed literature. Real silicon proves something works. It doesn't yet prove it's better than the alternative.
Our take
Our assessment: there's solid evidence that agents noticeably speed up design, verification and optimization at the RTL and FPGA level. A layperson can now get an entire language model running bit-exactly on an FPGA, and OpenAI's changes work in silicon. What hasn't been shown is that AI-invented architectures beat human designs in real hardware. Frostbite's 4.59x and Redwood's 3.4x per watt exist only in computational models. The figures measured on actual FPGA boards are more modest.
Three questions help when you read announcements like these. Which rung of the ladder is the project on? What's the comparison: its own first version, an unnamed core, an equally priced machine? And was it measured or projected? The next real tests will come from Architect Labs with its promised TSMC tape-out and from AMD with the first measured chip containing AI-made changes.
Sources
- Ankit Gupta: Vibecoding an inference chip ankitg.me
- OpenAI: Advancing chip design with our models openai.com
- Architect Labs: Introducing Frostbite architectlabs.com
- GitHub: SigmanticAI/apex-inference-chip github.com
- arXiv: Design Conductor 2.0 arxiv.org
- arXiv (HTML): Design Conductor 2.0 arxiv.org
- GeekNews: 추론 칩을 바이브 코딩하기 news.hada.io
- explainx.ai: Frostbite – claimed vs. verified explainx.ai
- DEV Community: What Frostbite's 4.59x claim actually rests on dev.to
- Notebookcheck: AI designed an all-new CPU notebookcheck.net
- VentureBeat: Architect Labs has to prove Redwood in silicon venturebeat.com
- PyTorch Korea: APEX FPGA LLM discuss.pytorch.kr
- HPCwire: Verkor unveils VerTQ hpcwire.com
- Tom's Hardware: The state of agentic AI in chip design tools in 2026 tomshardware.com
- Wikipedia: AlphaChip (controversy) en.wikipedia.org
- Google DeepMind: How AlphaChip transformed computer chip design deepmind.google
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.