SiMa.ai raises $150 million to scale the developer tool it pitches against Nvidia's CUDA

Take a perception engineer at a delivery-drone maker whose model already runs on an Nvidia Jetson module. SiliconANGLE puts the mainstream Jetson T2000 and T3000 at around 400 and 865 TOPS, kept to a lower power tier to preserve battery life ^[16]. A cheaper, lower-power chip is an easy sell to the finance team. The engineer's first question is how long it takes to get the existing model running on the new part.

SiMa.ai's answer is Palette Neat, which SiliconANGLE calls the company's "industry-first agentic development environment for Physical AI" ^[8]. The round also funds the next generation of Modalix, SiMa.ai's system-on-chip for embedded robotics ^[4]. According to SiliconANGLE, the offer is custom silicon plus software support that can deploy in days or hours, set against a CUDA architecture the outlet calls among the most complex and power-hungry ^[20].

What teams tell themselves users do is compare TOPS per watt and price, then switch. What users on an existing toolchain actually do, in my view, is stay put until a port is cheap enough to try without a project plan. Palette Neat is SiMa.ai's bid for that second group.

Krishna Rangasayee, the founder and chief executive and a former chief operating officer of Groq ^[9], pitches the whole stack. "While others are still figuring out the pieces or repurposing their cloud offerings, we've built the entire puzzle," he said ^[6]. The competitive case in both reports still leans on price and power. Rangasayee said the money would help "catapult" SiMa.ai into a dominant position against Nvidia, which SiliconANGLE describes as making more expensive CUDA-based hardware ^[7]. TechCrunch framed the bet around low latency and more affordable chips compared to Nvidia's GPUs ^[10]. Neither report includes a developer count for Palette Neat or a measured porting time on a customer's model.

SiMa.ai's 2028 target ^[5] sits about 16 percent above the T3000 and at half the 2,000 TOPS that SiliconANGLE lists for Nvidia's top Jetson AGX Thor modules ^[15][17][18]. The Thor figure is quoted at FP4 precision, while SiMa.ai's is quoted as plain tera operations per second ^[5][15]. SiMa.ai wants the new part to take it into higher-end drones, humanoid robots, driver-assistance systems and AI cockpits ^[11]. Thor is the module SiliconANGLE describes as designed for full-scale humanoids ^[15].

Investors have paid up before the 2028 part exists. The new valuation is about 1.5 times the Series B mark ^[1][19], and total funding now tops $500 million, according to TechCrunch ^[13].

Two facts about a device maker's own product settle whether any of this applies: whether it is power-bound in the 400 to 865 TOPS tier, and how much of its stack is custom CUDA code. Where the product is power-bound and carries little custom CUDA, a pilot is cheap and the price and power case applies directly. Deep custom CUDA in a power-bound product puts the decision on Palette Neat, and the days-or-hours claim is worth testing on the hardest model in the stack before any price talk. With power headroom and a thin CUDA layer, SiMa.ai is mostly leverage in a Jetson price negotiation. Teams with deep CUDA and no power pressure, or a need for Thor-class compute before 2028, get nothing from this round yet.