top of page
Business & Enterprise
AI Technology
Conversational AI / AI Assistant
Baseten
Built a high-growth AI inference business by making production model deployment fast, reliable and cost-efficient across clouds, then monetizing GPU utilization and API consumption despite hyperscalers’ infrastructure scale and model providers’ direct APIs.
1
MODEL
BUSINESS MODEL
SaaS
- Founded in 2019 by Tuhin Srivastava, Philip Howes, Amir Haghighat and Pankaj Gupta.
- Built a serverless inference platform for deploying, scaling and managing AI models in production.
- Rented GPU capacity from multiple cloud providers rather than operating its own data centers.
- Created a multi-cloud control plane for autoscaling, model routing, observability and performance optimization.
- Monetized open-model API consumption and dedicated GPU capacity by the minute or hour.
- Expanded from inference into multi-node training with H100 and B200 GPUs in 2025.
1. Identify a technical bottleneck that becomes more expensive as AI adoption grows.
2. Build an orchestration layer that improves performance, reliability or cost without owning the underlying infrastructure.
3. Price on the resource or outcome that scales with customer production usage.
4. Use multiple infrastructure suppliers to reduce capacity and pricing dependence.
5. Add adjacent workloads only when the same orchestration capability creates measurable value.
Avoid: building a thin wrapper over one cloud provider; if the provider can offer the same capability natively, the orchestration layer has weak pricing power.
DISTRIBUTION MODEL
Baseten scaled without owning data centers by orchestrating inference across clouds including AWS, Google Cloud and Microsoft Azure. Revenue grew sixfold in the fiscal year ending January 2025, then more than tenfold in the following 12 months; by 2026, estimated annualized revenue reached about $600 million, while the company processed more than one billion inference calls daily.
HOW THEY BUILT IT
HOW TO ARCHITECT IT
Developer/API-Led, Product-Led Growth, Direct Sales, Partnerships
HOW THEY OPERATIONALIZED
HOW TO REPLICATE WHAT WORKED
dm
- ML engineers discovered Baseten through documentation, open-source model deployment guides and developer communities.
- Self-serve deployment and serverless APIs let teams test models without infrastructure procurement.
- Direct sales pursued AI-native companies and enterprises with production inference workloads.
- Enterprise deployments allowed customers to bring their own infrastructure, expanding into security-sensitive accounts.
- Customers such as Cursor, Notion, Abridge and OpenEvidence created credibility with AI-native builders.
What worked: Let technical users deploy a real workload quickly, then expand into enterprise accounts once production usage proves performance and reliability.
Trap if copied blindly: Developer-led distribution fails when the product requires deep infrastructure integration or when the buyer needs procurement, security and compliance support before testing; the cost is long sales cycles and low self-serve conversion.
| PATTERNS OF THIS MODEL
PATTERNS IN MULTI-CLOUD AI INFRASTRUCTURE:
1. Orchestration creates value above raw compute. Customers pay for utilization, latency, reliability and cost efficiency, not merely GPU access.
2. Serverless entry reduces adoption friction. Developers can deploy without provisioning clusters or negotiating infrastructure contracts.
3. Dedicated capacity captures enterprise scale. High-volume customers move from shared APIs to reserved GPU deployments with predictable performance.
THE WEAKNESS: Multi-cloud orchestration has thin structural protection if cloud providers improve native inference services or if model providers capture the customer relationship.
What companies with this model reveal
| OPPORTUNITY INTELLIGENCE
GOLDMINE 1 — PERFORMANCE-BASED INFERENCE PRICING. Customers care about latency, throughput and cost per successful task; pricing tied to service-level objectives could capture more value than GPU minutes alone.
GOLDMINE 2 — AI RELIABILITY AS A PRODUCT. Production AI needs autoscaling, failover, observability and cost controls; reliability tooling can become a premium enterprise layer.
GOLDMINE 3 — MODEL-ROUTING ECONOMICS. Routing requests across models and hardware based on cost, latency and quality can create savings that customers will share with the platform.
THE PIT — GPU MARGIN COMPRESSION. If Baseten resells cloud GPU capacity, cloud price increases or demand spikes can compress gross margin.
THE SECOND PIT — HYPERSCALER DISINTERMEDIATION. AWS, Google Cloud and Azure can bundle inference optimization directly into their platforms.
MOVE WITH CAUTION: Do not expand into training infrastructure without proving that orchestration creates durable advantage; training is capital-intensive and places Baseten in more direct competition with hyperscalers.
Untapped Business Model / Gaps / Goldmines / Pits
Patterns & Insights
2
MARKET
mkt mt es
Baseten competed against hyperscalers, Modal, Replicate, Together AI and model providers’ native inference APIs. Rather than trying to outspend AWS, Google Cloud and Microsoft on data centers, it built a multi-cloud inference layer that optimized deployment, autoscaling and GPU utilization.
MARKET TYPE
Emerging Market
ENTRY STRATEGY
Greenfield Entry
WHY THEY WON
AI inference infrastructure was an emerging market shaped by rapidly changing models, GPU scarcity and uncertain enterprise deployment patterns. Baseten exploited a structural opening: companies wanted production AI without managing GPU clusters or committing to one cloud. Its positioning changed the buying criterion from “which cloud has GPUs?” to “which platform delivers the best performance, reliability and cost per inference?”
EXECUTION
Baseten created a new inference-orchestration layer rather than replacing an incumbent hosting product or reselling one cloud. Its critical bet was that AI teams would pay for faster deployment and better GPU utilization if Baseten removed infrastructure operations. This was viable because it could rent capacity from multiple clouds and sell performance improvements; the approach fails when cloud providers offer equivalent native capabilities or when customers already have mature internal ML platforms.
FOOTHOLD STRATEGY
Wedge Strategy
Baseten first gained traction through a wedge of ML engineers and AI-native startups that needed to deploy custom or open-source models without managing Kubernetes, GPUs and scaling infrastructure. These users were reachable through developer communities, could start with serverless deployment, and expanded into dedicated GPU capacity as production traffic grew.
GROWTH CAMPAIGN
CAMPAIGNS THAT WORKED
SO WHAT: Baseten’s strongest commercial message was not “we have GPUs”; it was “we make inference faster and cheaper across clouds.”
KEY LEARNING: Infrastructure companies win when they translate technical orchestration into a customer economic outcome.
HOW TO USE: If selling infrastructure, quantify the customer’s cost, latency or reliability improvement rather than selling access to the underlying resource.
KEY LEARNING
THE STANDARD: In an emerging infrastructure market, the winner owns the orchestration layer that turns scarce resources into dependable customer outcomes.
RULE 1 — SOLVE A SCALING BOTTLENECK: GPU supply, latency and cost are more urgent than another dashboard.
RULE 2 — USE MULTI-SUPPLIER LEVERAGE: Access to multiple clouds reduces capacity risk and strengthens the value proposition.
RULE 3 — MOVE UP FROM SELF-SERVE TO DEDICATED CAPACITY: Production workloads need predictable performance and enterprise controls.
RULE 4 — KNOW WHEN ENTRY FAILS: If the hyperscaler can bundle the same orchestration natively, the standalone layer needs a decisive performance or neutrality advantage.
MARKET TYPE: Emerging Market (category).
gc
Market Context
| MARKET INTELLIGENCE
PATTERNS IN TECHNICAL SELF-SERVE INFRASTRUCTURE GROWTH:
1. Documentation drives qualified adoption. ML engineers search for deployment, scaling and optimization guidance before contacting sales.
2. Production usage creates expansion. A successful model deployment leads to more models, regions, traffic and dedicated capacity.
3. Performance proof displaces feature lists. Infrastructure buyers respond to latency, uptime, GPU utilization and cost per inference.
4. THE WEAKNESS: Technical self-serve can attract experiments that never reach production, producing compute costs without durable revenue.
| MARKET ENTRY PLAYBOOK
THE STANDARD: Greenfield entry is appropriate when no incumbent owns the connecting orchestration layer and customers face a new operational bottleneck.
RULE 1 — TARGET TECHNICAL EARLY ADOPTERS: ML engineers can validate performance and become internal advocates.
RULE 2 — MAKE THE FIRST DEPLOYMENT SELF-SERVE: Reduce time from model to production traffic.
RULE 3 — PROVE ECONOMIC VALUE: Show lower cost per inference, better latency or higher GPU utilization.
RULE 4 — REJECT WHEN THE BOTTLENECK IS CAPITAL, NOT SOFTWARE: If the customer mainly needs owned GPUs or data-center capacity, an orchestration layer alone will not win.
How to enter
| FOOTHOLD STRATEGY PLAYBOOK
THE STANDARD: A wedge strategy succeeds when one technical workflow creates adoption and expands into broader infrastructure needs.
RULE 1 — START WITH CUSTOM-MODEL DEPLOYMENT: Teams with fine-tuned or proprietary models have stronger reasons to avoid generic APIs.
RULE 2 — USE SERVERLESS ACCESS: Let customers begin without capacity commitments.
RULE 3 — EXPAND INTO DEDICATED INFRASTRUCTURE: High-volume customers will pay for reserved GPUs and predictable performance.
RULE 4 — AVOID PREMATURE TRAINING EXPANSION: Training requires different economics and deeper infrastructure commitment than inference.
How to get the first strong position
MARKET PATTERNS & PLAYBOOK
3
MONEY
money rev pri
Baseten chose GPU time and API consumption as its pricing meters because both rise with production AI usage. Serverless customers pay for model API consumption, while dedicated deployments pay by GPU-minute or GPU-hour, with rates varying by hardware from T4 to B200.
REVENUE MODEL
Usage-Based, Subscription
PRICING MODEL
Usage-Based Pricing, Tiered Pricing, Custom Enterprise Pricing
WHY THEY WON
Customers pay for inference API consumption on hosted open-source models or for dedicated GPU capacity by minute or hour. Revenue expands with inference volume, model deployments, GPU tier, regions and enterprise commitments; gross margin depends on Baseten’s ability to improve utilization and negotiate infrastructure costs.
Baseten prices serverless inference by API usage and dedicated deployments by GPU time. Published rates range from about $0.01052 per minute on a T4 GPU to about $0.16633 per minute on a B200, while idle serverless replicas scale to zero; enterprise deployments are quote-based.
TARGET AUDIENCE
AI-native startups (Serverless/API; want fast deployment without infrastructure operations); Enterprise ML teams (Dedicated GPU; want performance, security and scale); Application developers (Open-model APIs; want lower-cost inference than closed-model APIs); Regulated enterprises (Bring-your-own-cloud; want control over data and infrastructure).
CUSTOMER BUYING BEHAVIOUR
An ML engineer tests Baseten by deploying a model or calling an API, compares latency, throughput and cost, then moves the workload into production. As traffic grows, the company adds dedicated GPUs, multiple regions or enterprise controls through a sales-led expansion.
| PRICING INTELLIGENCE
What makes this model effective & make customers pay
AI teams are not buying GPUs; they are buying reliable production inference at an acceptable cost per outcome.
RULE 1 — PRICE THE RESOURCE THAT SCALES WITH SUCCESS: GPU minutes and API calls rise as customer applications gain usage.
RULE 2 — MAKE SERVERLESS ENTRY FRICTIONLESS: Let customers deploy before asking for infrastructure commitments.
RULE 3 — USE DEDICATED CAPACITY FOR SCALE: High-volume customers will pay for predictable latency and reserved performance.
RULE 4 — WATCH CLOUD-COST PASS-THROUGH: If GPU costs rise faster than pricing, revenue growth can destroy margin.
The buyer is purchasing production reliability and inference economics, proven by GPU utilization, inference volume and enterprise commitments.
PRICE & REVENUE
| Revenue Risk - The biggest threat to revenue stability
Cloud-cost pass-through: GPU price increases or supply constraints can compress Baseten’s margin because it rents infrastructure from clouds.
Hyperscaler disintermediation: AWS, Google Cloud and Microsoft can bundle inference optimization, autoscaling and GPU capacity natively.
Model-provider disintermediation: OpenAI, Anthropic and Google can capture customers through direct APIs before they need custom-model infrastructure.
Customer concentration: Large AI-native customers can drive disproportionate inference volume and revenue.
Capacity risk: GPU shortages or regional constraints can limit growth during demand spikes.
Training expansion risk: Multi-node training increases capital intensity and direct competition with hyperscalers.
Baseten raised a $150 million Series D at a $2.15 billion valuation in September 2025 and later raised $1.5 billion at an approximately $13 billion valuation in 2026; Sacra estimates about $600 million annualized revenue in March 2026, but Baseten does not publicly disclose audited ARR.
Where the model can break
4
MOTION
motion ge cs
Baseten runs a hybrid motion: ML engineers self-serve through serverless deployment and APIs, while enterprise sales handles dedicated GPU capacity, security requirements, custom infrastructure and large production workloads.
GROWTH EXPANSION MODEL
Platform Expansion, Land & Expand
HOW THEY EXPAND
Baseten expanded from serverless model deployment into dedicated GPU capacity, multi-cloud management, enterprise bring-your-own-cloud deployments and training infrastructure. Each move increased workload depth: inference created adoption, dedicated capacity captured scale, and training expanded the platform’s role in the AI lifecycle.
COMPETITIVE STRATEGY
Differentiation
HOW THEY COMPETE
Against hyperscalers, Baseten flanked with multi-cloud neutrality and inference-specific orchestration. Against model providers, it targeted teams deploying custom, fine-tuned or open-source models that needed control over infrastructure, latency and cost.
GROWTH ENGINE
GTM
Content Flywheel
An ML engineer deploys a model → production traffic increases → Baseten autoscales and optimizes GPU utilization → the customer adds models, regions or dedicated capacity → enterprise sales expands the account. The loop compounds when production workloads create switching friction and more inference volume.
Baseten runs a developer-led, enterprise-expanded GTM. Self-serve deployment acquires ML teams; usage-based billing converts adoption into revenue; enterprise sales expands accounts through dedicated GPUs, security requirements and multi-cloud deployments. The primary motion is product-led growth, with direct sales becoming critical for high-volume and regulated customers. The transferable implication is that infrastructure platforms need a fast technical entry point and a credible enterprise expansion path.
SUSTAINING MOATS
High Customer Lock-In
moat
The moat is inference orchestration plus production dependence: customers build deployment pipelines, model configurations, observability and scaling policies around Baseten, making migration operationally costly.
| MOAT INTELLIGENCE
THE STANDARD: AI-infrastructure moats last when orchestration improves customer economics and becomes embedded in production systems.
RULE 1 — PRODUCTION WORKLOADS CREATE FRICTION: Deployments, autoscaling rules and monitoring become part of the customer’s operating stack.
RULE 2 — MULTI-CLOUD ORCHESTRATION CREATES LEVERAGE: Access to multiple infrastructure providers reduces single-vendor capacity risk.
RULE 3 — PERFORMANCE PROOF CREATES PRICING POWER: Lower latency or cost per inference gives customers a reason to consolidate workloads.
THE SIGNAL: More than one billion daily inference calls, customers including Cursor, Notion, Abridge and OpenEvidence, and rapid revenue growth indicate production adoption; the moat does not protect against native hyperscaler inference services or a major GPU-cost shift.
Why this company remains defensible
ARR & TAKEAWAY
ARR Journey - what to do at each stage
PRE-$1M ARR — PROVE INFERENCE ORCHESTRATION. After its 2019 founding, Baseten needed to prove that ML teams would use a third-party platform to deploy production models.
$1–5M ARR — WIN AI-NATIVE EARLY ADOPTERS. Serverless deployment and API access created the first repeatable technical adoption.
$5–10M ARR — EXPAND INTO PRODUCTION WORKLOADS. Autoscaling, reliability and observability became essential as customer traffic grew.
$10–50M ARR — ADD DEDICATED GPU CAPACITY. High-volume customers moved from shared inference to reserved performance.
$50–100M ARR — BUILD MULTI-CLOUD LEVERAGE. Baseten expanded across AWS, Google Cloud, Microsoft Azure and other infrastructure providers.
$100M+ ARR — SCALE INFERENCE AND ENTER TRAINING. Revenue grew sixfold in the year ending January 2025, then more than tenfold over the following 12 months; Baseten launched multi-node training on H100 and B200 GPUs in 2025.
Rule: Prove a production bottleneck, monetize its usage, then expand into adjacent AI workloads.
COPY PLAYBOOK : What Worked → What Failed → What to Replicate → What to Avoid
THE STANDARD: Build an orchestration layer that improves the economics of a scarce technical resource, then expand from self-serve adoption into enterprise-scale deployments.
SEQUENCE:
1. First, identify a resource bottleneck that grows with customer success; prerequisite: customers already feel the cost or complexity.
2. Next, make the first deployment self-serve and measurable; gate: technical users can prove performance or cost improvement quickly.
3. Then, add enterprise capacity, controls and adjacent workloads only after production usage is proven; gate: the expansion deepens the same customer workflow without weakening gross margin.
WORKED: Baseten used serverless inference to reduce deployment friction, then used multi-cloud GPU orchestration to capture production scale.
CAUTION: Do not build on one cloud provider without a neutrality advantage. Also avoid entering training before inference economics and enterprise demand are proven; training increases capital intensity and competitive exposure.
bottom of page