In 2018, Krishna Rangasayee left a senior role at a chipmaker to build something the cloud giants had written off: silicon that runs AI on the device itself, not in a distant data center.
The skepticism was technical. A vision model inside a drone or a robot arm has to run on a few watts. The usual answer — a general-purpose GPU — either drains the battery or cooks the chassis. SiMa.ai's bet was a system-on-chip designed for inference first, wrapped in software that shortens deployment from months to days.
On September 28, that bet got a public price. SiMa.ai closed a $150 million Series C at a $1.45 billion valuation, co-led by Fidelity Management & Research Company and Amplify.
The $1.45 billion mark prices physics, not sentiment: power and latency beat raw throughput in machines that move.
The open risk is timing. Design wins in automotive and robotics take years to convert, while NVIDIA keeps compressing the price of mid-range inference.
Oversubscribed round co-led by Fidelity and Amplify
Bringing total capital raised to $500 million. · SiMa.ai, 2026
Valued at $1.45B after three rounds
Founded in 2018. Institutions joined as new investors. · TechCrunch, 2026
The meter on every inference
Cloud inference has a meter. Every token and every frame is billed, and the bill scales with usage. For a chat product, that is tolerable. For a fleet of 10,000 machines, it is a permanent tax on operations.
Edge inference inverts the model. You pay upfront — silicon, integration, certification — and then the marginal cost of one more inference is close to zero.
That inversion is the whole investment case. A robot that reasons locally keeps working when the network drops. It does not stream video to a data center. It does not pay per call.
SiMa.ai sells this as a full stack: purpose-built silicon plus a software layer it calls Palette Neat, described by the company as the first agentic environment for physical AI. The framing is aggressively packaged.
While others are still figuring out the pieces or repurposing their cloud offerings, we've built the entire puzzle.— Krishna Rangasayee, founder and CEO, SiMa.ai
The verifiable part is simpler. The company says the platform compresses deployment from months to days. That is the claim product teams will test.
As we wrote in September, the edge-AI silicon race has moved from lab papers to commercial shipping. The question is no longer whether inference leaves the cloud. It is who captures the margin when it does.
What half a billion dollars buys
Total capital raised is now $500 million. The investor list reads less like a venture syndicate and more like a crossover book.
| Parameter | Detail |
|---|---|
| Round | $150M Series C, oversubscribed |
| Co-leads | Fidelity Management & Research Company; Amplify |
| New investors | AllianceBernstein, Baron Capital, J.P. Morgan, State of Michigan |
| Participating | Alter Venture Partners, Dell Technologies Capital, Maverick Capital, +ND Capital, Point72, StepStone Group |
| Use of funds | Scale Palette Neat; fund next-generation silicon |
Round structure. Source: SiMa.ai, September 2026
The money has two jobs. The first is commercial: push Palette Neat across robotics, automotive, drones, industrial automation, and healthcare. The company names Bosch, Emerson, STIGA, Synopsys, and Virya Autonomous Technologies among its customers and partners.
The second job is silicon. SiMa.ai says its next-generation hardware will deliver 1,000 dense TOPS, spanning machine-learning IP, chiplets, and systems-on-chip, slated for the first half of 2028.
That 2028 date is the real clock. It defines the window in which today's design wins either convert into volume or quietly expire.
Why institutional money is entering edge silicon now
A crowded lane
SiMa.ai is not alone in chasing inference at the edge.
Hailo sells vision processors into cameras and industrial systems. BrainChip has pushed neuromorphic silicon toward commercial deployment. Qualcomm bundles edge AI into the same mobile-class platforms that already sit inside cars and robots. On the data-center side, challengers attack the very layer SiMa.ai wants machines to stop renting.
The crowding matters for two reasons.
It validates the demand. When several well-funded teams, and the incumbents, converge on the same bottleneck, the bottleneck is real.
It also caps pricing power. Edge silicon is a design-win market: once a chip is designed into a platform, it tends to stay for the product's life. That stickiness is the prize. It is also why the fight happens years before revenue appears.
SiMa.ai's differentiator is the software layer. Selling silicon alone invites a feature-by-feature comparison against larger suppliers. Selling a deployment environment shifts the comparison from TOPS per watt to time-to-production, and that is a race a smaller team can win.
That logic explains the $500 million. The company is not raising to build one more chip. It is raising to become the default path from model to shipped device — and to hold that position until the 2028 hardware cycle decides the category.
Where the edge thesis can break
The bull case assumes the bottleneck stays where SiMa.ai points it: power, latency, and the cost of moving data.
Two things can move it.
First, NVIDIA. The incumbent keeps pushing smaller, cheaper inference parts into the same sockets. If a general-purpose chip gets good enough at low power, the specialist premium compresses.
Second, demand timing. Robots and vehicles are slow to certify. A design win today can sit in a pilot for three years before it ships in volume — and revenue follows shipments, not announcements.
The market backdrop is genuinely large. Counterpoint Research projects cumulative shipments of physical-AI devices, spanning robotics, automotive, and drones, at 145 million units by 2035. Large markets reward the winner and starve everyone else.
What would have to be true for this round to look cheap?
Probability: 45% — the technology fits, but incumbents and certification timelines both have to cooperate.
✅ Arguments for
Confirmation criteria: a named high-volume design win with shipping volumes and a second hardware generation that hits the promised 1,000 TOPS.
❌ Arguments against
Falsification criteria: a 2028 slip on the next-generation part, or a flagship customer switching to a general-purpose accelerator.
Shipping volumes on named humanoid and ADAS platforms, not pilot counts.
Whether the 2028 part arrives on schedule with 1,000 dense TOPS.
NVIDIA's pricing on mid-range inference silicon.
Follow-on rounds at a higher mark — or down rounds across edge-AI peers.
Scenarios
🟢 Bull case (35%)
Consequence: the $1.45 billion valuation looks conservative by 2029.
🟡 Base case (50%)
Consequence: a good business priced like a great one.
🔴 Bear case (15%)
Consequence: the round marks the top of the category's private valuation.