Lambda is a GPU cloud provider that rents NVIDIA GPUs on-demand to AI researchers, startups, and enterprises for training and inference workloads.[1] Founded in 2012 as a facial recognition startup by twin brothers Stephen and Michael Balaban,[2] the company pivoted to GPU infrastructure in 2017-2018 and has since grown into one of the leading "GPU-first" cloud platforms, branded as the "Superintelligence Cloud."[3]
Lambda's core value proposition is simple: pure GPU rental with zero egress fees, fast provisioning, and InfiniBand networking.[4] Unlike Crusoe or CoreWeave, Lambda does not own data centers or energy assets. It operates a capital-light model, leasing colocation capacity from partners like Aligned, Cologix, and EdgeConneX.[5] This makes Lambda fundamentally a GPU aggregator and cloud orchestration layer, not a vertically integrated infrastructure player.
Lambda is not building managed inference — they deprecated their Inference API and Lambda Chat assistant in September 2025.[10] Their strategic direction is pure GPU rental at scale. This makes Lambda a potential GPU supply partner for the platform, not a direct competitor on inference-as-a-service. The platform should monitor for any re-entry into inference APIs, but the current trajectory is clear: Lambda sells GPUs by the hour, the platform sells intelligence by the token.
| Name | Title | Background |
|---|---|---|
| Stephen Balaban | Co-Founder & CEO[2] | University of Michigan (CS & Economics). First engineering hire at Perceptio (acquired by Apple).[2] |
| Michael Balaban | Co-Founder & CTO[2] | University of Michigan (Math & CS). Twin brother of Stephen. Technical architecture lead.[2] |
| Peter Seibold | CFO[8] | Financial strategy and IPO preparation |
| Paul Miltenberger | VP of Finance[8] | Financial operations |
| David Hall | VP of NVIDIA Solutions[8] | Key NVIDIA relationship management |
| Robert Brooks IV | VP of Revenue[8] | Sales and go-to-market |
| Ariel Nissan | General Counsel[8] | Legal and compliance |
Lambda's leadership team is founder-led and engineering-heavy. There is no SVP of Product, no Head of Inference, no CPO. This confirms their strategic focus on infrastructure rather than managed AI services. Compare this to Crusoe, which has an SVP of Product and is actively hiring 8+ PMs including a Staff PM for Managed Inference.
| Round | Date | Amount | Valuation | Lead Investors |
|---|---|---|---|---|
| Seed (multiple)[13] | 2015-2018 | ~$4M | -- | Gradient Ventures, 1517 Fund, Bloomberg Beta |
| Series A[13] | Jul 2021 | $15M (+$9.5M debt) | -- | -- |
| Series B[13] | Mar 2023 | $44M | -- | Mercato Partners |
| Series C[13] | Feb 2024 | $320M | $1.5B | US Innovative Technology (Thomas Tull), B Capital, SK Telecom, T. Rowe Price |
| Series D[14] | Feb 2025 | $480M | $4B+ | Andra Capital, SGW, NVIDIA, ARK Invest, In-Q-Tel |
| Series E[6] | Nov 2025 | $1.5B+ | $5.9B | TWG Global (Thomas Tull & Mark Walter), USIT |
| Pre-IPO (planned)[9] | 2026 | ~$350M | TBD | Mubadala Capital (in discussions) |
| Total Raised | ~$2.3B+ |
| Year | Revenue | YoY Growth | Customers | Key Driver |
|---|---|---|---|---|
| 2022 | ~$20M (est.) | -- | ~2,000 | Hardware sales + early cloud |
| 2023 | $250M[7] | ~1,150% | 5,000+ | Cloud GPU rental explosion (ChatGPT-driven demand) |
| 2024 | $425M[7] | 70% | 10,000+ | Enterprise + AI lab adoption |
| 2025 (annualized) | $505M+[7] | ~19% | 10,000+ | NVIDIA leaseback + Microsoft deal |
Lambda's revenue growth decelerated sharply from ~1,150% (2023) to 70% (2024) to ~19% (H1 2025 annualized). This is consistent with GPU cloud pricing compression industry-wide. The NVIDIA leaseback and Microsoft deals provide revenue backstops, but Lambda will need to demonstrate renewed acceleration for a successful IPO. For the platform: This confirms that raw GPU rental is commoditizing. Managed inference margins will hold better.
NVIDIA signed a $1.5B deal to lease back 18,000 of its own GPUs from Lambda. The deal consists of two components: a $1.3B four-year lease for 10,000 GPUs and a $200M arrangement for 8,000 additional processors. This makes NVIDIA Lambda's single largest customer and provides predictable, locked-in revenue ahead of the IPO.
Lambda will deploy tens of thousands of NVIDIA GPUs, including GB300 NVL72 systems, in Lambda's liquid-cooled U.S. data centers for Microsoft's AI workloads. The deal cements Lambda as infrastructure-for-hire to hyperscalers, not a competitor to them.
Lambda's business model is increasingly that of a GPU broker and infrastructure lessor to hyperscalers and NVIDIA itself. This is a fundamentally different business from The platform's inference-as-a-service. Lambda sells capacity; The platform sells capability. They could be complementary.
Lambda offers four product tiers: self-serve GPU cloud instances, 1-Click Clusters for large-scale training, a Private Cloud for dedicated enterprise environments, and Echelon for on-premise deployments.[1]
Lambda's product stack has a conspicuous gap at the top: no managed inference layer, no model serving, no API gateway, no KV-cache optimization. Their inference API deprecation in September 2025 was a deliberate strategic retreat to focus on raw compute.[10] This is the exact layer The platform is building with the inference platform. Lambda sells the GPUs; the platform could buy them and sell intelligence on top.
| Specification | Detail |
|---|---|
| GPU Options | NVIDIA HGX B200 SXM6, H100 SXM[4] |
| Cluster Size | 16 to 512 GPUs per cluster[4] |
| GPU Interconnect | NVIDIA Quantum-2 400 Gb/s InfiniBand, rail-optimized topology[4] |
| RDMA Bandwidth | Up to 3,200 Gb/s per node (GPUDirect RDMA)[4] |
| Ethernet | 2x100 Gb/s per node (IP communication)[4] |
| Direct Internet | 2x100 Gb/s DIA connections per cluster[4] |
| Management Nodes | 3x CPU head nodes (jump boxes, Slurm scheduling)[4] |
| Software | Lambda Stack (PyTorch, TF, CUDA, cuDNN preloaded)[2] |
| SHARP Acceleration | Yes (InfiniBand in-network computing)[4] |
| Egress Fees | $0 (zero egress, zero ingress)[4] |
| Specification | Detail |
|---|---|
| Scale | Up to 165,000 NVIDIA GPUs per cluster[3] |
| GPU Types | GB200/GB300 NVL72, HGX B200, H100[3] |
| Tenancy | Single-tenant, dedicated infrastructure[3] |
| Networking | Quantum-class InfiniBand[3] |
| Cooling | Liquid-cooled racks[3] |
| Use Case | Large-scale AI training (frontier models) |
| GPU | Memory | On-Demand ($/hr) | Notes |
|---|---|---|---|
| NVIDIA HGX B200 | 192 GB HBM3e | $4.99[17] | 2x VRAM of H100, 3x faster training, 15x faster inference |
| 1-Click Cluster (B200) | 192 GB | $4.49/GPU[4] | 16-512 GPUs, InfiniBand included |
| NVIDIA H100 SXM | 80 GB | ~$2.49 | Pricing varies; mid-range for market |
| NVIDIA A100 SXM | 80 GB | ~$1.10[17] | Legacy GPU, still popular for inference |
| NVIDIA GH200 | 96 GB | Contact sales | Grace Hopper combined CPU+GPU |
| Feature | Lambda | Crusoe | CoreWeave | AWS |
|---|---|---|---|---|
| H100 On-Demand | ~$2.49/hr | $3.90/hr | $3.99/hr | ~$5.00+/hr |
| A100 On-Demand | ~$1.10/hr | $1.95/hr | ~$2.21/hr | ~$3.90/hr |
| Egress Fees | $0 | $0 | Yes | $0.09/GB |
| Managed Inference | No | Yes | Yes | Yes |
| Spot Instances | No | Yes | Yes | Yes |
| Multi-Chip Support | NVIDIA only | NVIDIA + AMD | NVIDIA only | NVIDIA + Trainium |
Lambda's aggressive pricing ($1.10/hr for A100) reflects the commoditization of raw GPU rental. H100 pricing has fallen from $8/hr in early 2023 to under $3/hr across most GPU clouds.[17] This compression makes The platform's strategy of selling managed inference (per-token pricing with higher margins) the right call. Raw compute is a race to the bottom; intelligence is not.
| Segment | Notable Customers | Relationship Type |
|---|---|---|
| Hyperscaler | Microsoft[12] | Multibillion-dollar infrastructure deal. Lambda hosts NVIDIA GPUs for Microsoft AI workloads. |
| GPU Manufacturer | NVIDIA[11] | $1.5B leaseback: NVIDIA leases back 18,000 of its own GPUs from Lambda over 4 years. Also an investor. |
| AI Labs | OpenAI, xAI, Anthropic[12] | Cloud compute customers for training and inference workloads. |
| Enterprise | Apple, Amazon, Google, Tencent[2] | Hardware sales (Echelon, workstations) and cloud GPU rental. |
| Defense / Gov | Raytheon, DoD[2] | On-prem Echelon deployments. In-Q-Tel is an investor.[14] |
| Research | MIT, Stanford, top universities[2] | Academic pricing, Echelon clusters. |
| AI Startups | 150,000+ cloud users[12] | Self-serve GPU cloud instances. |
Lambda's funnel is instructive: self-serve creates volume, clusters create stickiness, enterprise deals create revenue. The platform should study this progression. The self-serve tier is the acquisition engine; the managed inference tier (which Lambda abandoned) is where The platform can capture value that Lambda leaves on the table.
Lambda's two largest customers (NVIDIA and Microsoft) likely represent 50%+ of forward revenue. The $1.5B NVIDIA leaseback[11] and multibillion-dollar Microsoft deal[12] dwarf the self-serve cloud revenue. This creates customer concentration risk that IPO investors will scrutinize.
Lambda's customer base includes 150K+ AI developers and 10K+ paying customers who need GPU compute for inference. These customers are currently renting raw GPUs and managing their own inference stacks. The platform's managed inference platform could serve this exact audience with a turnkey solution at lower total cost of ownership.
| Location | Partner | Capacity | Status | Details |
|---|---|---|---|---|
| Kansas City, MO[15] | Owned (former bank DC) | 24MW (to 100MW) | Launching Early 2026 | $500M investment. 10K+ Blackwell Ultra GPUs. Single customer, multi-year. |
| Dallas-Fort Worth, TX[5] | Aligned | Liquid-cooled | Operational | High-density AI racks. |
| Columbus, OH[5] | Cologix | Multi-rack | Operational | HGX B200 clusters. "Days not weeks" deployment.[5] |
| Chicago, IL[5] | EdgeConneX | 23MW (build-to-density) | RFS 2026 | Single-tenant. Ready for GB300 NVL72. |
| Atlanta, GA[5] | EdgeConneX | Air-cooled | Operational | Two sites (ATL02). |
| San Francisco, CA[5] | Cologix (ECL) | Multi-rack | Operational | Lambda HQ region. |
| Allen, TX[5] | Cologix | Multi-rack | Operational | |
| Los Angeles, CA[5] | Prime (LAX01) | Multi-rack | Operational | |
| Additional Sites | Various | -- | Operational | 15+ total data centers across the US.[5] |
Lambda's infrastructure model is fundamentally different from Crusoe (vertically integrated) or hyperscalers (owned campuses). Lambda leases colocation space from third-party data center operators and deploys its own GPU racks within those facilities.[5]
| DC Ownership | Mostly leased (1 owned in KC)[15] |
| Energy Ownership | None |
| Capex Model | Capital-light (GPU procurement + lease) |
| Speed to Deploy | Fast (plug into partner DCs) |
| Margin Risk | Higher (pass-through power + colo costs) |
| Scale Target | 3 GW, 1M+ GPUs (aspirational)[5] |
| DC Ownership | Owned (modular containers) |
| Energy Ownership | Yes (structural cost advantage) |
| Capex Model | Capital-intensive but higher margins |
| Speed to Deploy | Modular (air-cooled containers) |
| Margin Risk | Lower (energy costs locked in) |
| Scale Target | Sovereign-ready, multi-chip |
Lambda's leased model means they face pass-through margin pressure on power and colocation costs. When GPU pricing compresses (as it is doing), Lambda's margins get squeezed from both sides: lower GPU rental prices AND fixed infrastructure costs. The platform's energy ownership provides a structural cost advantage on the largest operating expense line (power). This is The platform's moat.
| Metric | Lambda | CoreWeave | Crusoe | Nebius |
|---|---|---|---|---|
| Valuation | $5.9B[6] | $65B (public)[18] | $10B+[18] | $24.3B (public)[18] |
| 2024 Revenue | $425M[7] | $1.9B[18] | ~$276M[18] | ~$240M[18] |
| Employees | ~700[8] | 1,500+[18] | 1,000+[18] | 2,000+[18] |
| Total Funding | $2.3B+[6] | $12.9B+[18] | $3.9B[18] | $2.6B+[18] |
| Managed Inference | No (Deprecated)[10] | Yes | Yes (MemoryAlloy) | Yes |
| Own Data Centers | Mostly No[5] | Limited | Yes | Yes |
| Own Energy | No | No | Yes (45 GW) | No |
| Multi-Chip | NVIDIA only | NVIDIA only | NVIDIA + AMD | NVIDIA only |
| Egress Fees | $0[4] | Yes | $0 | Yes |
| Key Differentiator | 1-Click Clusters, zero egress | NVIDIA early access, K8s-native | Vertical integration, MemoryAlloy | Yandex heritage, EU presence |
| Anchor Customer | NVIDIA, Microsoft[11][12] | Microsoft | OpenAI/Oracle | EU enterprises |
Lambda occupies a unique position in the GPU cloud landscape: it is the most developer-friendly of the GPU-first clouds (self-serve, zero egress, simple pricing), but it is also the least vertically integrated (no owned DCs, no energy, no managed services). This makes Lambda best-suited for:
Lambda is least suited for:
SemiAnalysis ClusterMAX 2.0 ratings place Lambda in the Silver tier, behind CoreWeave (Platinum) and Crusoe (Gold).[19] The rating reflects Lambda's weaker networking, storage, and orchestration capabilities compared to peers who have invested more heavily in managed platform services.
Lambda is not a direct threat to The platform's inference-as-a-service strategy for three clear reasons:
| # | Reason | Evidence |
|---|---|---|
| 1 | Lambda exited managed inference | Deprecated Inference API and Lambda Chat in Sep 2025.[10] No product team building inference services. |
| 2 | Lambda sells raw compute, not intelligence | Their entire product is GPU-hours. The platform sells tokens. Different customers, different value prop. |
| 3 | Lambda is NVIDIA-only, not multi-chip | Cannot offer workload-optimal routing across GPU vendors. A multi-chip architecture is a genuine differentiator. |
| # | Opportunity | Details |
|---|---|---|
| 1 | GPU supply partnership | Lambda has 15+ US data centers with NVIDIA GPUs. the platform could lease GPU capacity from Lambda for burst inference workloads while building out its own infrastructure. |
| 2 | Managed inference overlay | Lambda's 150K+ users need inference but Lambda stopped offering it. the platform could offer a managed inference layer on Lambda GPUs. |
| 3 | Echelon + the platform inference stack | Lambda Echelon serves on-prem enterprise and government customers. The platform's inference software could run on Echelon hardware. |
| # | Risk | Trigger to Watch | Likelihood |
|---|---|---|---|
| 1 | Lambda re-enters inference | New inference API announcement, hiring of inference engineering team, or product leadership (CPO/SVP Product) | Medium |
| 2 | Post-IPO capital deployment | Lambda IPOs in H2 2026 and uses proceeds to build managed services layer[9] | Medium |
| 3 | Lambda acquires inference company | Acquisition of a Together AI, Fireworks, or similar inference platform | Low |
| 4 | Lambda builds sovereign offering | In-Q-Tel investment[14] suggests government interest; could lead to classified/sovereign cloud | Low |
Lambda has excess GPU capacity across 15+ US data centers. the platform could lease burst capacity from Lambda while building its own infrastructure. Lambda's zero egress policy makes this operationally simple.
Lambda's 150K+ users lost their inference API in Sep 2025. The platform's the inference platform can serve this audience with managed inference at lower TCO than self-managed GPU rental.
Lambda's IPO (H2 2026) will generate significant capital. Watch for inference product announcements, product leadership hires, or acquisitions that signal re-entry into managed services.
Both Lambda and Crusoe offer zero egress fees. This is becoming the market standard for GPU-first clouds. the platform must match this or clearly justify any data transfer charges.
| Feature | Details |
|---|---|
| Scale | Single rack (40 GPUs) to data center scale (1,000s GPUs)[16] |
| Compute Nodes | Lambda Hyperplane servers: 4, 8, or 16 NVIDIA GPUs per node with NVLink[16] |
| Networking | InfiniBand HDR 200 Gb/s or 100 Gb/s Ethernet[16] |
| Storage | Proprietary and open-source options[16] |
| Software | Lambda Stack: PyTorch, TensorFlow, CUDA, cuDNN (preloaded, regularly updated)[16] |
| Support | Premium/Max tiers with direct phone access to AI infrastructure engineers[16] |
| Customers | Fortune 500, top universities, DoD[16] |
Lambda's executive team has no product management leadership, no Head of Inference, no Chief Product Officer, and no DevRel leader. The VP-level roles are focused on finance, NVIDIA relationships, and sales. This organizational structure confirms that Lambda is an infrastructure and sales organization, not a product-led platform company. For the platform, this means Lambda is unlikely to build competitive managed services in the near term.
| Investor | Type | Round(s) | Significance |
|---|---|---|---|
| TWG Global (Thomas Tull, Mark Walter)[6] | Mega-fund ($40B AUM) | Series E Lead | $1.5B single check. Largest AI infrastructure bet. |
| NVIDIA[14] | Strategic | Series D + Leaseback | Investor AND $1.5B customer. Dual relationship. |
| US Innovative Technology Fund[13] | Thomas Tull vehicle | Series C, E | Repeat lead investor across rounds. |
| In-Q-Tel (CIA)[14] | Government | Series D | US intelligence community interest in Lambda infrastructure. |
| ARK Invest[14] | Public market crossover | Series D | Cathie Wood's fund. Pre-IPO positioning. |
| Mubadala Capital[9] | Sovereign wealth (Abu Dhabi) | Pre-IPO (in talks) | Potential $350M convertible note lead. |
| T. Rowe Price, SK Telecom[13] | Institutional + Strategic | Series C | Blue-chip institutional validation. |
This report was compiled from 24 primary sources including Lambda's corporate website, product documentation, press releases, SEC-related filings, investor announcements, third-party research (Contrary Research, Sacra, SemiAnalysis), and industry publications (TechCrunch, CNBC, Tom's Hardware, Data Center Dynamics, BusinessWire). Revenue figures are sourced from Sacra Research and Contrary Research estimates. Organizational structure is inferred from public executive profiles. All data accessed February 14-16, 2026.
This report was compiled from 24 primary sources including Lambda's corporate website, product documentation, press releases, third-party research (Contrary Research, Sacra, SemiAnalysis), investor announcements, and industry publications (CNBC, TechCrunch, Tom's Hardware, Data Center Dynamics, BusinessWire). Revenue projections are sourced from Sacra Research. Organizational structure is inferred from public executive profiles (Clay, Craft.co, LinkedIn). All performance claims are from Lambda's own documentation unless otherwise noted. Report accessed and compiled February 14-16, 2026.