Score Labs emerged from stealth with Score Engine, inference software running open-weight AI models on standard Intel Xeon servers, shown at Microsoft Ignite.
Key Takeaways
- Score Labs claims Score Engine runs 30% to 4x faster than vLLM and SGLang on x86 servers.
- Intel-validated tests showed 1.5x better cost efficiency than an NVIDIA H100 GPU.
- Funding, investors and valuation were not disclosed.
Lead
Score Labs came out of stealth this week with Score Engine, enterprise inference software built to run open-weight AI models on standard Intel Xeon servers instead of scarce GPUs. The company is demonstrating the product at Microsoft Ignite 2026 on the Intel Booth Theater stage. CEO Sean Cook and co-founder and Chief AI Systems Architect Ashish Kumar present on Thursday, November 19, at 10:00 a.m. PT, and live demos run twice daily from November 17 to 20. Score Labs has not disclosed a funding round, lead investor or valuation.
What Does Score Engine Actually Do?
Score Engine is an inference stack that serves open-weight models on x86 CPUs, GPUs or a hybrid of both, with data staying on the customer's own hardware. Score Labs says it wrote six proprietary libraries from scratch. They control how a model uses a processor's math units, memory and network connections. The company has filed provisional patents on the work.
The software competes with open-source serving engines such as vLLM and SGLang, which most teams run on GPUs. It also competes with hosted APIs, which many enterprises use because they lack data-center GPU capacity of their own. Score Labs pitches a third option: reuse the Xeon servers already in the rack.
How Fast and How Cheap Is It?
Score Labs reports 30% to 4x more performance than vLLM and SGLang, a wide range that depends heavily on workload. The range matters more than the headline, because a 30% gain is a tuning story and a 4x gain is an architecture story.
The strongest number carries Intel's name. In testing validated by Intel on long-running AI agent tasks, servers with four Intel Xeon 6 or 5th Gen Xeon processors delivered 1.5x better cost efficiency than one NVIDIA H100. That is a cost-per-unit-of-work comparison, not a raw speed claim, and it covers one class of task. Score Labs also says its approach costs 3x to 17x less than hosted APIs, benchmarked against OpenRouter.
One customer result is public. A dental AI platform replaced a commercial AI service with Score Engine, cutting projected monthly inference cost by more than 80% and getting better results. The customer is unnamed, and the figure is a projection.
Which Workloads Is It Targeting?
Score Labs is aiming at regulated, document-heavy work where data residency matters: healthcare, legal, financial services and software development. The Ignite demos cover legal due diligence, anti-money-laundering transaction review, freight and logistics document processing, and coding agents. Each server handles 8 to 16 simultaneous coding sessions. It also shows multimodal tasks, including text-to-speech, video summarization and image generation.
These are mostly throughput-tolerant jobs. A review of 500 contracts overnight does not need the sub-100-millisecond latency that a consumer chatbot does, and that is where CPU inference makes the most sense.
Why Does the Intel Xeon Angle Matter Now?
The Xeon angle matters because GPU supply and pricing still shape enterprise AI budgets, and Intel has an obvious interest in showing that its server CPUs can carry inference. Intel placed Score Labs on its own booth stage at a Microsoft event, and Microsoft is listed alongside Intel as a partner. For a company that has just left stealth, that is useful distribution. It also means the most prominent benchmark comes from a vendor with a stake in the result.
The skeptical reading is straightforward. CPUs have long trailed GPUs on memory bandwidth, the main limit on large-model token generation. A software gain of 4x over open-source engines does not close that hardware gap for every model size. The cost-efficiency win against an H100 depends on workload shape, batch size and how the servers are priced.
What Comes Next for Score Labs?
The next test is independent replication and named customers. Intel-validated numbers and one anonymous dental platform are an opening position, not a track record. Buyers will want results on their own models, and Score Labs will need to show how it handles rapid releases of open-weight models, since each new architecture can require fresh low-level tuning.
The funding question is also open. A stealth exit at a major conference usually precedes a capital raise or a disclosure of one, but Score Labs has said nothing on either. Details on headquarters and headcount were also absent from the announcement.
Outlook
Score Labs enters a crowded inference market with a narrow, testable claim: standard Xeon servers can run enterprise AI at lower cost than a GPU for certain workloads. The Ignite demos and Intel's validation give it early credibility, though customer proof remains thin. The next signals to watch are named deployments, independent benchmarks and any financing announcement.


