Tsavorite Raises $5M Seed From Pavestone to Scale 'Omni Processing Unit' AI Chip With $100M+ in Pre-Orders (May 4, 2026)
Bengaluru–Silicon Valley AI chip startup Tsavorite, founded by Intel veterans, raised a $5M seed round led by Pavestone VC on May 4, 2026 — adding fresh capital to a company already sitting on more than $100M in pre-orders for its Omni Processing Unit, a single-chip device that fuses CPU, GPU and memory.
AI hardware startup Tsavorite Scalable Intelligence announced on that it has closed a $5 million seed round led by Pavestone VC, adding fresh capital to a company already carrying more than $100 million in pre-orders for its forthcoming Omni Processing Unit (OPU) — a single chip that the founders claim consolidates CPU, GPU and memory while cutting AI inference energy use by roughly 90 percent.
What Happened
Tsavorite, founded in 2023 by five Intel veterans — CEO Shalesh Thusoo (previously at Marvell, Tanzanite Silicon Solutions, Intel and Cisco), Supriya Madan, Guntram Wolski, Sarvagya Kochak and Shirish Seetharam — first emerged in November 2025 with its OPU architecture and a public claim of more than $100 million in pre-orders from Fortune Global 500 companies, sovereign cloud providers and systems integrators.
The new $5 million round, confirmed by Pavestone VC and reported by YourStory, Inc42 and Entrackr, is described as a strategic seed to scale the team and bring the OPU to first silicon. Tsavorite operates across the United States and India, with a substantial design centre in Bengaluru handling much of the chip and software work.
Key Details
- Round size: $5 million, seed stage, led by Pavestone VC
- Pre-orders: More than $100 million already booked, per the company's own disclosures
- Product: Omni Processing Unit (OPU) — unifies CPU, GPU, memory and scale-up/scale-out connectivity in a single device
- Performance claims: 10× faster AI processing at roughly 10 percent of the energy use of current GPUs
- Headline IP: MultiPlexus fabric — an in-house die-to-rack interconnect with petabyte-scale bandwidth and gigabyte-scale caches
- Delivery: First silicon and Helix-class AI appliances targeted for shipment in 2026
- Founding team: Five Intel veterans, design centre in Bengaluru, headquarters in the United States
Why It Matters for Developers and AI Teams
Tsavorite is the latest entrant in a fast-growing class of startups arguing that the only way to break the GPU-and-DRAM bottleneck is to redesign the silicon. Competitors include UK-based Fractile, which is in early talks with Anthropic about its DRAM-less SRAM-based architecture, plus Cerebras, Tenstorrent and Groq. The case for unified-memory chips is concrete: large-language-model inference is increasingly memory-bandwidth bound, not compute bound, and existing GPU stacks pay a heavy round-trip tax shuffling weights between HBM and compute units.
The OPU's design pitch — CPU, GPU and memory on the same die, connected by a custom MultiPlexus fabric — is one of several attempts to remove that tax. If Tsavorite ships silicon on schedule and the performance numbers hold, it would give AI inference operators a cheaper alternative to Nvidia H200/B200 racks for serving open-weight models.
What Developers and Investors Are Saying
The reaction across EE Times, DCD and Indian startup press has been measured. Analysts are taking the $100 million pre-order figure seriously because the founders have shipped real silicon at Intel and Marvell before, and Pavestone's involvement — a fund with deep semiconductor relationships — lends credibility. Skeptical commenters note that pre-orders are not revenue, that no independent benchmarks have been published, and that "CPU+GPU+memory on one die" has been promised before by chip startups that ultimately missed their tape-out targets.
The decision to base the design centre in Bengaluru while keeping the headquarters in the US has been called out as a deliberate cost play. Founder Shalesh Thusoo has said the team explicitly evaluated RISC-V and ultimately picked Arm cores, citing the 20-plus years it has taken Arm to mature server-class software support.
What's Next
Tsavorite has said it expects to deliver first silicon and Helix-class enterprise AI appliances in 2026, capable of running "all agentic AI workflows with leadership performance." Watch for an updated benchmark disclosure, the first independent third-party performance numbers, and confirmed customer deployments — the company's $100 million pre-order book is a strong signal but only as good as the silicon that ships against it.
Sources
- YourStory — primary funding announcement
- Inc42 — coverage of the seed round and full-stack AI platform plans
- Entrackr — founder backgrounds and Indian operations
- Data Center Dynamics — pre-order disclosure and OPU architecture
- EE Times — Arm vs. RISC-V architectural choice
- Inside HPC — original OPU launch coverage from November 2025
Stay up to date with Doolpa
Subscribe to Newsletter →