Alex Cheema is co-founder and chief executive of EXO Labs, which builds software for running powerful AI models on hardware people own. With co-founder Mohamed Baioumy, he created EXO, an open-source system that combines ordinary computers into inference clusters, and launched local.ai to help people navigate local-AI hardware, models, and benchmarking.
Cheema co-founded a game-automation startup at 16, studied physics at the University of Oxford, and served as chief technology officer of GamingForGood before building decentralized-finance infrastructure at Portofino. His earlier projects include Graphite, for constructing succinct proofs about graphs, and Tanuki, a trustless wallet-scoring protocol. In 2024, he and Baioumy co-authored an AI and cryptography primer examining verifiable inference, decentralized computing, federated learning, and payments for AI agents.
The technical ideas behind EXO
- Distributed inference across consumer hardware. EXO automatically discovers devices, maps their connections, and distributes models across their combined memory. Cheema has demonstrated open-weight models running across interconnected Mac Studios and explored mixed Apple Silicon–NVIDIA configurations. RDMA over Thunderbolt 5 reduces communication latency between Macs, making tensor parallelism more practical when every generated token requires repeated synchronization.
- Prefill/decode disaggregation. Processing an incoming prompt is typically compute-bound; generating subsequent tokens is often constrained by memory bandwidth. Cheema describes pairing an NVIDIA DGX Spark with a MacBook: the Spark handles prefill, the Mac handles decoding, and the key-value cache moves between them. The arrangement works only when network overhead does not consume the gains.
- Intelligence per joule. Cheema evaluates local systems by useful model capability relative to energy consumption, not processor specifications alone. Memory capacity, bandwidth, caching, kernel efficiency, and application harnesses all influence whether inference remains fast, affordable, and practical outside a data center.
- Benchmarks must preserve model quality. Aggressive quantization can inflate apparent performance while damaging the answers a model produces. Through local.ai, Cheema frames deployment as choosing both suitable hardware and the right combination of model, quantization, inference engine, configuration, and harness. His MacBook-versus-FPGA benchmark similarly tests specialized-hardware claims against measured results.
Cheema sees user-controlled personal AI as a foundation for privacy, continuity, and independence from hosted providers. His longer-term interests include persistent agents, multi-agent workloads, test-time compute, and individualized models. He remains clear-eyed about the obstacles: large downloads, expensive memory, awkward networking, and software that still demands too much expertise.
