SambaNova Systems is a custom silicon AI startup founded in 2017 by Stanford professors Kunle Olukotun and Christopher Ré, along with former Oracle SVP Rodrigo Liang.[1] The company built the Reconfigurable Dataflow Unit (RDU), a purpose-built AI processor that takes a fundamentally different architectural approach from NVIDIA GPUs.[2] After raising $1.14B and achieving a $5B peak valuation in 2021, SambaNova struggled to convert technical differentiation into sustainable commercial traction.[3]
By late 2025, Intel entered advanced acquisition talks at $1.6B including debt, a 68% decline from peak valuation.[4] Those talks stalled in January 2026, and SambaNova pivoted to raising $350M+ from Vista Equity Partners and Intel in a new Series E round.[5] The company has laid off 15% of its workforce and pivoted from training to inference-only.[6]
SambaNova is a cautionary tale, not a competitive threat. Their trajectory from $5B to $1.6B validates The platform's core thesis: betting on a single custom silicon architecture creates vendor lock-in risk that enterprise buyers increasingly reject. The platform's GPU-agnostic, multi-chip approach (NVIDIA, alternative silicon) is the correct strategic posture. SambaNova's talent pool and distressed IP represent potential acqui-hire and partnership opportunities rather than competitive risks.
SambaNova was born from Stanford University's Pervasive Parallelism Laboratory, where co-founders Kunle Olukotun (known as the "father of the multi-core processor") and Christopher Ré (a MacArthur "Genius" Award recipient) developed the theoretical foundations for dataflow computing applied to AI workloads.[1] Rodrigo Liang, who had led nearly 1,000 chip designers at Oracle as SVP of Hardware Development, joined as CEO to commercialize the technology.[9]
The company name means "new dance" in Portuguese, reflecting Liang's vision of a fundamentally different approach to AI computation.[9]
| Name | Title | Background |
|---|---|---|
| Lip-Bu Tan | Executive Chairman[8] | CEO of Intel; former CEO of Cadence Design; appointed May 2024. Major conflict-of-interest concern. |
| Rodrigo Liang | Co-Founder & CEO[9] | 20+ years semiconductor engineering. SVP Hardware at Oracle/Sun Microsystems. Led ~1,000 chip designers. |
| Kunle Olukotun | Co-Founder & Chief Technologist[1] | Stanford Professor of EE/CS. "Father of the multi-core processor." Director of Stanford Pervasive Parallelism Lab. |
| Christopher Ré | Co-Founder[1] | Stanford Associate Professor of CS. MacArthur Fellow. Affiliated with Stanford AI Lab. |
| Amarjit Gill | Board Director[8] | Venture investor |
| Catriona Mary Fallon | Board Director[8] | Appointed August 2024 |
Intel CEO Lip-Bu Tan simultaneously serves as SambaNova's Executive Chairman. When Intel pursued a $1.6B acquisition of SambaNova in December 2025, Reuters reported that the deal would have directly boosted Tan's personal fortune.[8] This conflict of interest contributed to regulatory and board scrutiny, and the acquisition talks ultimately stalled. Tan also served on SoftBank's board until 2022; SoftBank is both an Intel investor and SambaNova's largest backer via Vision Fund 2.
| Round | Date | Amount | Valuation | Lead / Key Investors |
|---|---|---|---|---|
| Series A[3] | Mar 2018 | $56M | -- | Walden International, GV (Google Ventures) |
| Series B[10] | Apr 2019 | $150M | -- | Intel Capital, GV |
| Series C[3] | Feb 2020 | $250M | ~$2.5B | BlackRock, existing investors |
| Series D[3] | Apr 2021 | $676M | $5.0B | SoftBank Vision Fund 2, Temasek, GIC, BlackRock |
| Series E[5] | Feb 2026 | $350M+ | TBD | Vista Equity Partners, Intel ($100-150M) |
| Total | ~$1.5B+ |
| Investor | Stake | Strategic Interest |
|---|---|---|
| SoftBank Vision Fund 2 | Largest (led $676M Series D)[3] | AI hardware portfolio play. Massive unrealized loss at $1.6B vs. $5B entry. |
| Intel Capital | Early investor (Series B)[10] | Strategic: AI inference accelerator for Intel ecosystem. Lip-Bu Tan connection. |
| Vista Equity Partners | New lead (Series E)[5] | Enterprise software PE firm. Via partnership with Cambium Capital. |
| BlackRock | Series C, D participant[3] | Financial investor. Likely underwater on position. |
| Temasek / GIC | Series D participants[3] | Singapore sovereign wealth. AI infrastructure thesis. |
| GV (Google Ventures) | Series A, B[3] | Strategic hedge. Google also builds own TPUs internally. |
SambaNova's 68% valuation decline from peak ($5B in April 2021) to Intel's $1.6B offer (December 2025) is one of the largest value destructions among well-funded AI chip startups. For context:
SambaNova has never publicly disclosed revenue figures. The company's commercial strategy has shifted significantly over its history:
SambaNova's lack of disclosed revenue despite $1.14B in funding is a significant signal. It suggests that custom silicon companies face enormous difficulty in building recurring revenue businesses outside of government contracts. The platform's inference-as-a-service model, which is hardware-agnostic and API-first, avoids this trap.
SambaNova offers three main products, all powered by the SN40L RDU chip. The product lineup has evolved significantly as the company pivoted from hardware sales to managed services.[11]
Cloud-based inference API service. Developers access SambaNova's RDU-powered inference via a standard OpenAI-compatible API.[11] Supports DeepSeek R1 671B, Llama 3.x family, Qwen, and other open-source models. Claims to be the fastest inference platform for large models.
A "turnkey" managed inference cloud for data centers. SambaNova ships, installs, and manages SambaRack hardware in the customer's existing data center.[12] Claims 90-day deployment vs. the typical 18-24 month cycle. Targets enterprises and sovereign governments that need on-premise AI but lack hardware expertise.
Full-stack enterprise AI platform for customers who want to purchase and manage their own SambaNova hardware. Includes SambaStudio software, model management, and optimization tools. Previously the core product; now de-emphasized in favor of managed offerings.
SambaNova's most distinctive software innovation is their Composition of Experts model architecture.[13] Unlike traditional Mixture of Experts (MoE), CoE aggregates multiple small "expert" models (e.g., 150 x 7B models = 1 trillion parameter system) with a router model that directs queries to the optimal expert. The SN40L's three-tier memory enables 15-31x faster model switching vs. GPU-based approaches, because experts can be hot-loaded from DDR to HBM in microseconds rather than milliseconds.
The SN40L is SambaNova's fourth-generation Reconfigurable Dataflow Unit, manufactured on TSMC's 5nm process.[2] It represents a fundamentally different architectural approach from GPUs: instead of executing programs as sequences of kernel launches, the RDU maps entire neural network computation graphs directly into hardware as streaming dataflows.[14]
| Parameter | SN40L RDU | Comparator: NVIDIA H200 |
|---|---|---|
| Architecture | Streaming Dataflow[14] | GPU (CUDA cores) |
| Process Node | TSMC 5nm[2] | TSMC 4nm |
| Transistor Count | 102 billion[2] | 80 billion |
| Compute Units | 1,040 PCUs + 1,040 PMUs[14] | 16,896 CUDA + 528 Tensor Cores |
| Peak Compute | 638 BF16 TFLOPS[14] | 989 BF16 TFLOPS |
| On-Chip SRAM | 520 MiB (PMU SRAM)[14] | ~50 MB L2 Cache |
| HBM | 64 GiB HBM[14] | 141 GB HBM3e |
| DDR DRAM | Up to 1.5 TiB[14] | N/A (system RAM separate) |
| TDP | ~600W per chip[15] | 700W |
| Cooling | Air-cooled[15] | Liquid-cooled (typically) |
The SN40L's most significant architectural innovation is its three-tier memory hierarchy, which directly addresses the "memory wall" problem that limits GPU inference performance for large models.[14]
| Parameter | Specification |
|---|---|
| RDUs per Rack | 16 SN40L chips[15] |
| Aggregate Compute | 10.2 PFLOPS (BF16)[15] |
| Total HBM | 1,024 GiB (16 x 64 GiB) |
| Total DDR | Up to 24 TiB (16 x 1.5 TiB) |
| Interconnect | Peer-to-peer RDU network |
| Form Factor | Standard 19" rack, air-cooled[15] |
| Power Consumption | ~10 kW average[15] |
Unlike NVIDIA's latest Blackwell systems (GB200 NVL72) which require liquid cooling, SambaNova's SambaRack operates with standard air cooling in a 19" rack.[15] This significantly lowers deployment complexity and cost, and is directly relevant to The platform's air-cooled infrastructure strategy. A single SambaRack consuming ~10 kW can run DeepSeek R1 671B at 198 tokens/sec, whereas comparable GPU systems would require multiple racks and significantly more power.
SambaNova has made aggressive performance claims, particularly around inference speed and energy efficiency. Many of these are self-reported and should be validated independently. However, the Stanford HAI benchmarks provide third-party validation for some efficiency claims.[7]
| Model | Metric | SambaNova Claim | Context |
|---|---|---|---|
| DeepSeek R1 671B | Tokens/sec[16] | 198-255 tok/s | Full model (not distilled), 16 SN40L RDUs. "World record" per SambaNova. |
| Llama 3.1 405B | Tokens/sec[17] | 132 tok/s | First to claim real-time inference on full 405B model. |
| Llama 3.3 70B | Tokens/sec | 500+ tok/s | Per SambaCloud benchmarks. |
| CoE Routing | Model switch time[13] | Microseconds | 15-31x faster than GPU-based model loading. |
| Metric | SambaNova vs. NVIDIA H200 | Source |
|---|---|---|
| Intelligence per Joule | 4x better than Blackwell[7] | Stanford HAI benchmark (third-party) |
| Low Batch Performance | 9x faster, 5.6x more efficient[7] | SambaNova self-reported vs. H200 |
| High Batch Performance | 4x faster, 2.5x more efficient[7] | SambaNova self-reported vs. H200 |
| Chips Required (DeepSeek R1) | 16 RDUs vs. 300+ GPUs[16] | 95% fewer chips claim |
| Power per Rack | ~10 kW vs. 40-120 kW[15] | Air-cooled vs. liquid-cooled GPU racks |
| Company | Chip | Architecture | Key Claim | Status |
|---|---|---|---|---|
| SambaNova | SN40L RDU | Dataflow | 198 tok/s DeepSeek R1 671B | Shipping |
| Cerebras | WSE-3 | Wafer-scale | 1,800 tok/s Llama 3.1 70B | Shipping |
| Groq | LPU | TSP dataflow | ~500 tok/s Llama 70B | Shipping |
| Etched | Sohu | ASIC (transformer-only) | 500K tok/s Llama 70B (claimed) | Pre-production |
| NVIDIA | B200 | GPU | Broad ecosystem, dominant | Shipping |
SambaNova's customer base is concentrated in government/national labs and a small number of enterprise accounts. The company has struggled to achieve broad commercial adoption, which is a pattern common to custom silicon vendors.
| Customer | Sector | Deployment | Use Case |
|---|---|---|---|
| Argonne National Lab[18] | DOE / Government | SN40L inference cluster (16 RDUs) + legacy SN30 training system | AuroraGPT foundation model for scientific research (biology, chemistry, materials, climate) |
| Los Alamos National Lab[19] | DOE / Government | Expanded DataScale + SambaNova Suite | Generative AI and LLM workloads for national security research |
| Lawrence Livermore National Lab | DOE / Government | DataScale system | AI-driven scientific computing |
| Enterprise customers | Various | Undisclosed | SambaCloud API users (unknown volume) |
SambaNova's disclosed customers are overwhelmingly U.S. Department of Energy national laboratories. While these are prestigious deployments that validate technical capability, they represent a narrow, grant-funded customer base with limited commercial scalability. The company has not publicly disclosed any Fortune 500 enterprise customers or recurring revenue contracts. This government-lab concentration is a structural weakness.
SambaNova sold DataScale appliances directly to customers for $1M+ per system. This required long sales cycles, custom integration, and significant professional services. The model generated lumpy, non-recurring revenue.
Recognizing the limitations of hardware sales, SambaNova launched SambaCloud and SambaManaged to shift toward recurring revenue. SambaCloud offers free-tier access to attract developers (20 requests/min, 50/day), while SambaManaged targets enterprise data center operators with a 90-day deployment promise.[12]
| Tier | Access | Rate Limits | Credits |
|---|---|---|---|
| Free | API access, all models | 20 req/min, 50 req/day | None |
| Developer | API access, all models | 1,000 req/day (with $10 topup) | $5 starting + $50 bonus |
| Enterprise | Custom SLAs, dedicated | Custom | Custom pricing |
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| GPT-OSS 120B | $0.22 | $0.59 |
| Llama 3.1 405B[20] | $5.00 | $10.00 |
| DeepSeek R1 671B | Enterprise / Contact Sales | Enterprise / Contact Sales |
| Llama 3.3 70B | ~$0.60 | ~$0.60 |
SambaCloud's pricing is competitive but not disruptive. The free tier is useful for developer attraction but the 50 requests/day limit is restrictive. Enterprise pricing is opaque ("Contact Sales"), which signals that SambaNova is still figuring out its unit economics. Compare with The platform's potential to offer transparent, per-token pricing with guaranteed SLAs.
SambaNova operates in the custom AI silicon market alongside Cerebras, Groq, Etched, and others. All face the same existential challenge: convincing customers to adopt proprietary hardware in a market where NVIDIA's CUDA ecosystem is the de facto standard.
| Dimension | SambaNova | Cerebras | Groq | Etched |
|---|---|---|---|---|
| Valuation | ~$1.6B (implied)[4] | $7.6B (pre-IPO) | $2.8B | $3.3B |
| Chip Gen | 4th (SN40L) | 3rd (WSE-3) | 2nd (LPU) | 1st (Sohu) |
| Focus | Inference + CoE | Training + Inference | Inference speed | Transformer inference |
| Cooling | Air-cooled | Water-cooled | Air-cooled | TBD |
| Revenue | Undisclosed | $136M (FY2024) | Undisclosed | Pre-revenue |
| Key Risk | Valuation collapse | Customer concentration | Limited model support | Unproven hardware |
| Moat | 3-tier memory, CoE | Wafer-scale, training | Latency speed | Transformer ASIC |
SambaNova's trajectory illustrates the fundamental challenges facing all custom AI silicon companies:
The platform's GPU-agnostic, multi-chip inference platform is positioned to benefit from these dynamics rather than suffer from them. By abstracting the hardware layer and offering customers a single API that works across NVIDIA GPUs, and alternative silicon accelerators, the platform eliminates the vendor lock-in fear that constrains every custom silicon vendor. The platform can route workloads to the most cost-effective hardware for each query, delivering the performance benefits of custom silicon without the risk.
| Threat Vector | Level | Rationale |
|---|---|---|
| Direct competition (inference APIs) | LOW | SambaCloud has minimal commercial traction. Developer community is small. |
| Enterprise deals (SambaManaged) | MEDIUM | SambaManaged's 90-day deployment promise could appeal to sovereign/government customers. |
| Technology disruption | LOW | RDU is impressive but niche. No evidence of broad model coverage or ecosystem adoption. |
| Talent acquisition | OPPORTUNITY | Post-layoff engineers with dataflow and compiler expertise are available for recruiting. |
| Partnership opportunity | OPPORTUNITY | SambaNova hardware could be integrated into A multi-chip inference platform. |
SambaNova raises $350M+ and continues as an independent company focused on inference. They become a niche player serving government labs and select enterprises. Strategic implication: Evaluate SambaNova as a chip supplier for A multi-chip inference stack. Their air-cooled, energy-efficient hardware could reduce The platform's cost-per-token for specific workloads. Negotiate favorable supply terms while SambaNova needs commercial customers.
Acquisition talks resume and Intel buys SambaNova for $1.5-2B. RDU technology gets absorbed into Intel's data center chip portfolio. Strategic implication: Monitor closely. If Intel integrates RDU into its Gaudi/Xeon ecosystem, the platform may get access to dataflow inference through Intel's broader product line. Intel as a chip supplier is more reliable than standalone SambaNova.
Funding round fails or company runs out of runway. IP and talent become available at distressed prices. Strategic implication: Prepare an acqui-hire target list of 10-15 SambaNova engineers specializing in dataflow compilation, inference optimization, and chip-software co-design. Budget $2-5M for targeted recruitment. The RDU IP itself is likely too expensive to acquire directly.
SambaNova spent $1.14B trying to build both the chip and the platform. The platform can achieve better outcomes by building only the platform and leveraging multiple chip vendors (including potentially SambaNova) for the hardware layer. This is architecturally analogous to how AWS succeeded by abstracting hardware, while hardware-only companies struggled. SambaNova's distress creates an opportunity for the platform to access world-class inference silicon at favorable terms.
| Risk | Severity | Description |
|---|---|---|
| Valuation Death Spiral | HIGH | 68% decline from peak. Down rounds erode employee retention and customer confidence. |
| Revenue Opacity | HIGH | No disclosed revenue after $1.14B raised. Likely burning $100M+/year on chip R&D and operations. |
| Customer Concentration | HIGH | Visible customers are almost exclusively DOE national labs. Limited commercial enterprise traction. |
| Governance Conflict | MEDIUM | Intel CEO as Executive Chairman creates acquisition conflicts and regulatory scrutiny.[8] |
| Talent Retention | MEDIUM | 15% layoffs + valuation decline make it difficult to retain top engineers. Stock options likely underwater. |
| NVIDIA Roadmap | MEDIUM | Each NVIDIA generation (Blackwell, Rubin) narrows the performance gap with custom silicon. |
| Model Ecosystem | MEDIUM | Limited to open-source models (Llama, DeepSeek, Qwen). No proprietary model partnerships disclosed. |
| Manufacturing Risk | LOW | TSMC 5nm is mature and reliable. No known manufacturing issues. |
| Strength | Relevance to the platform |
|---|---|
| Energy efficiency (4x vs. Blackwell per Stanford) | Directly relevant to The platform's cost optimization. Lower power = lower inference cost. |
| Air-cooled deployment | Aligns with The platform's air-cooled infrastructure strategy. |
| Three-tier memory for large models | Enables running 671B models on 16 chips. the platform could leverage for large-model workloads. |
| Composition of Experts (CoE) | Novel routing architecture that could inform The platform's multi-model inference routing. |
| Government/lab relationships | The platform could partner or compete for sovereign inference contracts. |
| World-class technical team | Recruiting opportunity, especially post-layoff engineers. |
SambaNova is not a competitive threat to the platform. It is a potential chip supplier and talent pool. The company's trajectory from $5B peak to $1.6B Intel offer validates every assumption underlying The platform's GPU-agnostic inference platform strategy:
Bottom line: Watch SambaNova as a potential partner, recruit from their talent pool, and use their cautionary tale in every customer conversation about why The platform's approach is superior.