Sacra estimates that Wafer hit $8M in annualized revenue in August 2026.

Wafer reached this run rate within 12 weeks and reported gross margins of about 50% because it rents chips. Accelerator capacity from Nvidia and AMD accounts for most of its costs, so software optimizations increase the spread between what Wafer pays per GPU-hour and earns per token served.

Revenue comes from two channels: a usage-based serverless inference API priced per token and dedicated endpoints configured for a customer's model, traffic shape, and latency target. Named customers include Vercel and Inworld AI, while Wafer has also performed optimization work for DigitalOcean and Neon Health. The company has served more than 2 trillion tokens of continual inference.

Wafer was founded in San Francisco in May 2025 by CEO Emilio Andere and Steven Arellano. After participating in Y Combinator's Summer 2025 batch, it launched its commercial inference platform in mid-2026 and increased annualized revenue from effectively zero to $8M in roughly three months.

Wafer raised a $40M Series A co-led by Marathon Management Partners and Chemistry at a valuation above $200M, announced on September 1, 2026. Wing, AMD Ventures, Outset Capital, Fifty Years, and Y Combinator also participated. Angel investors included Jeff Dean, Guillermo Rauch, Andy Fang, Kyle Vogt, Akshay Kothari, Matthew Prince, Scott Stephenson, and Wojciech Zaremba.

Wafer previously announced a $4M seed round on April 14, 2026, led by Fifty Years with participation from Liquid2 and Y Combinator.

Total disclosed funding is approximately $44M. The company rejected multiple acquisition offers from larger inference and cloud providers before the Series A.

Wafer builds AI inference engineers that optimize open-source models for specific workloads and help customers run them. Its automated system profiles a customer's inference traffic, identifies bottlenecks across the serving stack, generates candidate improvements, tests them against model output, and deploys the highest-performing configuration.

Developers using the serverless product replace an OpenAI-compatible endpoint with Wafer's base URL and API key. Behind a standard chat-completion API, Wafer handles GPU provisioning, model loading, batching, scaling, and continuous retuning for open-weight models.

For dedicated inference, customers specify a model, expected traffic pattern, and service-level objective, such as sub-600ms time-to-first-token for a voice agent. Wafer benchmarks hardware and software configurations, provisions an isolated endpoint, and profiles production traffic for further optimization. Neon Health moved a voice-agent workload to a dedicated Wafer endpoint, reducing median time-to-first-token from roughly 800ms to 550ms at a higher peak load.

The optimization search covers serving engines such as vLLM and SGLang, quantization formats, speculative decoding, request scheduling, tensor and expert parallelism, custom GPU kernels, and hardware selection across Nvidia and AMD chips. Wafer plans to add support for accelerators from Google, SambaNova, Cerebras, and other non-Nvidia providers.

In July 2026, Wafer optimized Z.ai's GLM-5.2 model for AMD Instinct MI355X chips, reaching about 80% of Nvidia B200 performance at less than half the cost. In other published work, Wafer improved Kimi 2.5 throughput on AMD by 11.3x over a stock configuration and increased DeepSeek V3.2 single-request output speed from 38.5 to 200.8 tokens per second. The interventions varied by workload and included bug fixes in ROCm framework paths, custom kernels, topology changes, and scheduler tuning, indicating that inference performance depends on optimization across the stack rather than a single-layer fix.

Wafer is a software-differentiated inference provider with a blended B2B and B2C go-to-market. Serverless customers pay per token with no idle-capacity charges, while dedicated customers commit to reserved capacity with workload-specific SLAs. The company's earlier flat-rate subscription product, Wafer Pass, targeted individual coding-agent users, but Wafer now focuses primarily on serverless and dedicated enterprise inference.

Wafer's optimization software is designed to extract more usable tokens from each GPU-hour than a generic serving configuration. At roughly 50% gross margins, the company retains the spread between rented accelerator costs and per-token revenue. Improvements identified by the optimization agent, including faster kernels, better parallelism layouts, and routing workloads to cheaper AMD hardware, widen that spread without additional capital expenditure.

Hardware arbitrage provides another margin lever. AMD accelerators often offer more memory per dollar than Nvidia GPUs, but their ROCm software ecosystem is less mature. Wafer addresses that software gap through targeted kernel work and framework fixes, purchasing cheaper capacity while delivering competitive performance. The arbitrage may narrow as vendor software improves and competitors adopt similar techniques, making the speed and breadth of Wafer's continual optimization loop its primary defense.

The go-to-market follows a land-and-expand pattern. Developers test the serverless API by changing a base URL, then migrate to dedicated endpoints as usage grows and latency or compliance requirements tighten. Wafer also distributes through infrastructure partnerships, including DigitalOcean, and AI gateways such as TrueFoundry, where it can operate behind a routing layer without requiring each customer to integrate directly.

More customer traffic generates additional optimization traces, improving the agent's ability to identify valid speedups and lower cost per token. Lower costs can attract more traffic and generate further traces. The resulting dataset consists not of user text, but of performance traces, bottleneck patterns, candidate results, and correctness failures accumulated across workloads and hardware.

Wafer competes in a fragmented inference market. Direct rivals sell managed access to open-weight models, adjacent players address the same cost and latency problem through proprietary silicon, and hyperscalers bundle inference into broader cloud platforms.

Together AI is the closest full-platform competitor. Both companies offer OpenAI-compatible serverless inference, dedicated endpoints, and continuous optimization based on production traffic. Together's ATLAS speculative-decoding system learns from real workloads, directly overlapping with Wafer's continual-inference thesis. Together reports serving more than 400 trillion tokens per month and offers provisioned throughput, private VPC deployments, training, multimodal APIs, and multi-region failover.

Fireworks AI competes through a proprietary disaggregated inference engine that separates prefill and decode into independently scalable pools, as well as reinforcement fine-tuning, voice-agent infrastructure, and air-gapped enterprise deployments. Baseten targets custom-model and mission-critical inference with SOC 2 Type II certification, HIPAA support, and hybrid or self-hosted deployment options. DeepInfra, Replicate, fal.ai, RunPod, and Modal round out the direct competitive set, with different combinations of developer accessibility, GPU breadth, and infrastructure primitives.

Wafer differentiates through hardware neutrality and cross-stack optimization. Rather than requiring workloads to conform to one proprietary engine or accelerator vendor, Wafer searches across multiple engines, kernels, and chips for the configuration suited to each workload.

Cerebras, Groq, and SambaNova derive performance advantages from proprietary hardware rather than software optimization on commodity GPUs. Cerebras offers ultra-low-latency token generation from wafer-scale processors and recently partnered with AMD to combine rack-scale AMD systems with Cerebras engines. Groq's LPU architecture delivers fast decode. A December 2025 licensing agreement with Nvidia means LPU-style acceleration could become an incumbent feature bundled with Nvidia's broader platform.

SambaNova targets regulated enterprises with turnkey on-premises racks and has cited deployments at JPMorganChase. These players can win workloads where architectural specialization outperforms optimized GPU deployments, particularly for interactive decode latency on supported models.

Wafer can route workloads to the hardware with the best economics for a given model shape and traffic pattern rather than committing to one architecture. Its AMD work addresses the software gap on hardware with higher memory capacity and lower hourly cost.

AWS Bedrock, Google Vertex AI, and Microsoft Azure represent the largest structural threat. These platforms bundle inference into existing cloud contracts, IAM policies, networking, security reviews, and committed spend. AWS added fully managed open-weight models through Project Mantle in early 2026, with OpenAI-compatible endpoints and automated capacity management.

Nvidia NIM packages optimized model runtimes into containers deployable across cloud, data center, and edge environments, directly overlapping with Wafer's value proposition. Open-source serving engines such as vLLM and SGLang continually absorb techniques that were previously specialized, raising the unoptimized baseline and narrowing the proprietary advantage of managed inference providers.

Wafer argues that procurement convenience and generic optimization leave performance gains available for latency-sensitive and cost-sensitive workloads. Its published case studies report multi-fold improvements over stock configurations, but the company must find those gains faster than upstream frameworks and vendor libraries close the same gaps.

Wafer's expansion logic follows three vectors: broadening the hardware it optimizes, deepening its enterprise and partner channels, and extending its optimization platform beyond hosted inference.

Wafer runs workloads on Nvidia and AMD chips and plans to support accelerators from Google, SambaNova, Cerebras, and other non-Nvidia providers. Each additional architecture expands the optimization search space and the potential value of a neutral layer that benchmarks and routes workloads across heterogeneous hardware.

AMD Ventures' participation in the Series A gives Wafer access to AMD's roadmap, preferred GPU capacity, and joint go-to-market programs with AMD's cloud partners. As inference fragments across Nvidia, AMD, hyperscaler silicon, inference ASICs, and CPU-based deployments, hardware selection and workload routing become a larger part of the inference stack.

Wafer's DigitalOcean relationship uses a B2B2C model in which Wafer supplies model-level optimization while the cloud partner provides infrastructure, billing, and the customer relationship. Replicating this model with other clouds, AI gateways, and managed platforms would let Wafer distribute its technology without owning every end-customer relationship or financing GPU capacity in every geography.

Enterprise model usage is more fragmented than consumer usage, with adoption occurring through cloud platforms and open-source deployments rather than only OpenAI and Anthropic. Dedicated endpoints with SLAs, data residency, and compliance controls could extend Wafer into regulated verticals such as healthcare, where it already operates under a business associate agreement with Neon Health.

Wafer's largest TAM expansion vector is packaging its optimization agent as deployable software for customer VPCs, sovereign clouds, and on-premises infrastructure. This would shift the company from participating in inference spend to monetizing the underlying engineering layer, extending its addressable market to enterprises, AI labs, and chipmakers that must run workloads in their own environments.

Model-launch partnerships with open-weight labs could provide a recurring acquisition channel. Wafer could co-optimize kernels and serving configurations before release and operate official endpoints afterward, gaining early technical access when a new model generation ships. The proliferation of mixture-of-experts architectures, new attention mechanisms, and long-context techniques expands the search space, making generic serving stacks less likely to deliver optimal results without workload-specific tuning.

Optimization commoditization: Open-source serving engines such as vLLM and SGLang, Nvidia's TensorRT-LLM, and AMD's improving ROCm ecosystem continue to absorb previously specialized techniques, meaning a configuration that creates a multi-fold advantage today may become standard within months and requiring Wafer to compound proprietary optimization data and automation faster than those improvements become upstream features.

Margin compression from token-price deflation: Inference token prices have declined roughly 600-fold since 2020, with economy-tier price half-lives near one year, and if Wafer sells tokens at declining rates while GPU rental costs fall more slowly, efficiency gains may pass to customers rather than accrue to Wafer, making the company's 50% gross margin difficult to defend without a shift toward dedicated contracts, shared-savings pricing, or software licensing.

Customer and hardware concentration: With only a handful of named customers driving an $8M run rate and strategic investment from AMD Ventures, Wafer faces the risk that losing one large workload could leave specialized capacity underutilized, while competing chip vendors or potential cloud partners may view the company as aligned with one hardware ecosystem rather than as a neutral optimizer.