The Phoenix That Might Still Burn: Inside Mythic’s $125M Bet Against Nvidia
The Phoenix That Might Still Burn: Inside Mythic’s $125M Bet Against Nvidia
Mythic looked like it was running on fumes in 2023. A $13 million rescue round kept the lights on. Now the Palo Alto analog-chip startup says it’s back with a roar.
On Wednesday, Mythic announced a $125 million raise. The round was oversubscribed and DCVC led it. The company is also reviving a big promise: its analog processing units can use 100 times less energy than top GPUs on AI workloads. If that holds up, it would matter a lot for data centers that already feel like power-hungry furnaces.
The backers read like a flex. NEA is in. Future Ventures is in. SoftBank is in. Then it gets more revealing: Honda and Lockheed Martin also wrote strategic checks. That pairing doesn’t scream “replace Nvidia in hyperscale racks tomorrow.” It points to a nearer target where watts are life-or-death: perception systems in autonomous vehicles and military drones. In those worlds, “plug it into the wall” isn’t a plan.
Mythic also swapped leadership muscle into place. CEO Taner Ozcelik joined in 2024 after building Nvidia’s automotive business. Since arriving, he’s rebuilt the architecture and the software stack. Mythic’s only shipping chip, the M1076, puts up real numbers for edge inference: about 25 TOPS while sipping roughly 3–4 watts. That’s not a rounding error. It’s a legit achievement.
Still, the flashiest claims stay out on the horizon. Mythic talks about 750× better tokens-per-watt on trillion-parameter models. It also teases sub-1-watt sensors and pocket-sized LLMs. Those are projections for now. They’ll need independent proof before anyone should carve them into stone.
The Energy Crisis Is Real; the Fix Is Still a Maybe
You don’t have to be a grid engineer to see the squeeze coming. AI power draw isn’t a marketing phantom. It’s a hard constraint with a bill attached.
Some projections say data centers could hit 9–12% of U.S. electricity use by 2030. That’s up from about 4% in 2024. AI inference is a big driver of that growth. In other words, efficiency isn’t a virtue badge anymore. It’s a product feature.
Mythic’s pitch rests on analog compute-in-memory. The idea is to dodge the classic Von Neumann bottleneck. Traditional chips keep memory and compute separate, so they shuttle data back and forth. That “commuting” can burn roughly 90% of the energy in many setups. Mythic instead performs matrix multiplies directly inside analog memory arrays.
Here’s the basic bottleneck Mythic wants to avoid:
Memory ⇄ (data shuttling) ⇄ Compute Energy leaks in the commute. Mythic’s bet: do the math where the data lives.
On paper, the efficiency looks wild. Mythic cites about 17 femtojoules per multiply-accumulate operation. GPUs sit around 17 picojoules. That’s a 1,000× improvement at the transistor level. It’s the kind of number that makes investors sit up straight.
Then reality shows up with a clipboard. Analog-to-digital converters draw power. Calibration to fight noise and drift adds overhead. Many non-matrix operations still run digitally. When you measure the whole system, the gap tends to shrink.
So what’s actually been shown? Verified benchmarks point to about 3.8× power savings for the M1076 on specific workloads. That’s strong. It’s also nowhere near the headline “100×” framing.
And Mythic isn’t alone in this race. EnCharge AI raised over $100 million in February 2025 with similar efficiency promises. Rain Neuromorphics went looking for buyers after funding dried up. The space is crowded with companies offering order-of-magnitude leaps. In a market like that, the real moat often isn’t the slide-deck math. It’s manufacturing repeatability and software that developers don’t dread using.
What Venture Capitalists Actually Press For
Hardware dreams get tested by boring questions. The kind you ask when physics is the auditor.
Start with capacity. Mythic says the M1076 holds about 80 million weights per chip. Do the rough math on a trillion-parameter model. Even at one byte per weight, you’d need around 12,500 chips just to store the weights. That’s before you deal with activations, KV-cache, or redundancy. So if you hear “trillion-parameter datacenter inference,” you should also hear the unspoken footnote: Mythic’s story likely depends on undisclosed next-gen chiplets with far higher density.
This is the diligence checklist investors lean on in practice. Not as a slogan but as a demand for boundaries and receipts.
A quick view of “claims vs. what needs proving” helps:
| Topic | What Mythic is saying | What skeptics want to see |
|---|---|---|
| Tokens-per-watt | Up to 750× better on trillion-parameter models | Full system measurements, not just core array energy |
| Efficiency headline | 100× less energy than top GPUs | Independent benchmarks across real workloads |
| Scaling to datacenters | Big-model inference potential | Clarity on next-gen chiplets and weight density |
| Analog overhead | Manageable in practice | ADC power, calibration intervals, drift at steady-state |
| LLM practicality | Pocket-sized LLMs, sub-1-watt sensors | Where activations and KV-cache live, bandwidth costs included |
The sharpest questions keep coming back to one theme: system-level truth. Investors want tokens-per-watt with the measurement box drawn clearly. Include host processors. Include interconnects. Don’t cherry-pick the “core array” and call it the whole meal.
They also want MLPerf Inference results. They want calibration intervals disclosed. They want steady-state ADC overhead spelled out. And for LLM inference, they want a straight answer on activations and KV-cache. Memory bandwidth often dominates energy there, so “great math units” alone won’t save you.
Where Mythic’s Best Path Probably Runs
If you’re expecting Mythic to knock Nvidia off the datacenter throne next quarter, you’ll likely end up disappointed. Mythic’s most credible wedge sits in edge deployments where power is tight and latency needs to be deterministic. Offline operation matters there too. That’s exactly why design-in cycles, though slow, can be worth it.
Honda and Lockheed’s strategic bets fit that storyline. Those industries will pay extra for inference that works when you’re disconnected and thermally boxed in.
From here, three outcomes map the terrain.
The base case looks practical: Mythic becomes a meaningful edge-inference supplier. A defense prime or an automotive supplier could buy it once the fit looks proven.
The bull case is more ambitious: Mythic chiplets start offloading matrix ops alongside GPUs in hybrid systems. Think of it as a specialist engine bolted onto a bigger machine.
The bear case is the familiar hardware tragedy: calibration overhead stays stubborn. Software friction scares developers away. Wins come but they don’t repeat cleanly. Revenue turns lumpy and Mythic ends up niche.
One more thing stands out. Mythic has called GPUs “1945-era relics” and it’s framed this as a civilizational stake. That kind of rhetoric can sound like a company trying to justify a big round after a near-death wobble.
The power constraint is real. The open question is simpler: can analog deliver at scale when the whole system is counted? That’s what the $125 million is really buying time to answer.
NOT INVESTMENT ADVICE