Anna Goldie: the brain is still more efficient than AI
Ricursive CEO Anna Goldie told The Information the brain, at about 20 watts, is still more efficient than AI. Her thesis is chip design, not a brain chip.
The Ricursive Intelligence CEO told The Information that biology shows huge slack in energy and data. The fix, she says, is chip design.
On 28 September 2026, Anna Goldie appeared on AI Deep Dive episode 3, a The Information (TITV) program. The cofounder and CEO of Ricursive Intelligence argued that the brain is still more efficient than AI: today’s systems burn far too much energy and data to learn and to answer. The human brain, she said, runs on about 20 watts while handling perception, reasoning, and coordination at once. Heavy training and inference clusters take megawatts. The Information posted the clip on X.
That gap matters because the wait in AI is no longer just the model. It is the silicon, the power bill, and the one to three years until a chip built to run that model exists. Goldie did not ship a product that matches the brain. Treating that as a company result is RUMOR.
Quick recap
The comparison is watts and data volume, not consciousness. Transformers got as far as they did, in part, because they fit GPUs first built for graphics. Ricursive wants to shrink a chip design cycle from years to weeks. The end to end product still lacks a market proof.
Why the brain is still more efficient than AI
Efficiency here means watts and data volume. A desk lamp sits around 20 watts. A warehouse of servers sits in the megawatts. The real mechanism is different: a biological neuron is not a silicon transistor. The comparison fails if “brain” is read as a neural implant.
Ricursive works on electronic design automation (EDA), the software that places blocks and wires on a data center chip. It is not a brain-computer interface.
Transformers on the GPUs we already have
In the same episode, Goldie said Transformer models, the architecture behind most current chatbots, got as far as they did in part because they fit GPUs, chips first built for graphics. A GPU is a graphics processing unit; it now also trains large models because it is good at parallel math. The global optimum for AI architectures, in her view, is probably far from today’s silicon and neural nets. The path she described is codesign: the algorithm and the accelerator get designed together. In the official clip, the thesis is that “it’s all about the chips.”
Ricursive and AlphaChip
Goldie and Azalia Mirhoseini founded Ricursive Intelligence in December 2025. TechCrunch, TIME, and The Wall Street Journal report the same bet: shrink a chip design cycle that now takes one to three years down to weeks or days. The pair led AlphaChip at Google Brain and Google DeepMind. The method was published in Nature in 2021 as “A graph placement methodology for fast chip design.” Google DeepMind says it went into production on four generations of TPUs, Google’s in-house AI chips (v3, v4, v5e, and v5p/Trillium), and on the Axion data-center processor. A TPU is the accelerator Google builds for its own models.
How much the company raised
Ricursive has raised $335 million in total: a $35 million seed led by Sequoia Capital in late 2025 and a $300 million Series A led by Lightspeed Venture Partners at a $4 billion valuation in January and February 2026. DST Global, Felicis, and NVentures, Nvidia’s venture arm, are also in. Nvidia sells the accelerators that dominate training today. That stake is a conflict of interest.
On Sequoia’s Training Data podcast, the founders talk about moving from fabless (you design, someone else, often TSMC, fabricates) to what they call designless: AI draws the silicon from a software need. That is the company’s thesis, not an industry standard.
The Epoch AI institute estimates global AI compute has been doubling every seven months, a 3.3x annual rate in H100-equivalents between 2022 and 2026, a figure TIME also cited in its TIME100 AI 2026 list. An H100 is an Nvidia accelerator used as the comparison unit.
Promise versus reality
The company projects up to 10x better total cost of ownership (TCO, what you spend to own and run the chip over time) for accelerators built for specific models. That number is Ricursive’s own. There is no public independent test. Design partnerships are still early. The new software has no open commercial metrics.
The brain in the interview is proof that smart, energy-cheap computation is physically possible. It is not an announcement of parity with biological tissue.
This helps if you want to know why a chip-design startup is talking about brains. It does not help if you expected a neural implant, or a chip that already matches 20 watts.
What we still don’t know
There is still no market proof of the end to end product. We do not know whether the one to three year cycle drops to weeks outside Google’s lab, or whether the 10x TCO shows up in a chip that ships. This newsroom has not tested Ricursive’s software.
Goldie uses the brain to say there is still slack. Ricursive is selling a shorter cycle. The 20 watt gap remains an argument, not a product.
Sources
Interview
- The Information, AI Deep Dive episode 3
- The Information on X, clip of the remark
- The Information on X, clip on Transformers
Official and research
- Ricursive Intelligence
- Google DeepMind, AlphaChip
- Nature, 2021, A graph placement methodology for fast chip design
- PR Newswire, Ricursive Series A
- Sequoia Capital, essay on Ricursive
- Sequoia Capital, Training Data podcast
- Epoch AI, AI chip production
Press