AI is hitting a new gear. The biggest bottleneck isn’t raw computing power anymore. It’s memory. Yes, memory—the part of computer systems that stores and quickly retrieves data—is the choke point slowing down AI’s massive potential.

Two startups are taking bold swings to fix this. One builds chips that bring processing power closer to memory. The other reinvents how AI models reuse their own data during inference. Both just landed huge funding rounds from industry giants. This is the future of AI infrastructure, and it’s unfolding right now.

XCENA’s Memory-Adjacent Chip Shakes Up AI Compute

Meet XCENA, a four-year-old startup with a radical idea: cut the costly back-and-forth data trips between CPUs, GPUs, and memory. Normally, every AI request shuttles data out of memory, to processors, then back again. This wastes time and energy.

XCENA’s chip, called MX1, flips the script. It embeds thousands of small, specialized compute cores directly inside memory modules. This design lets the chip handle routine data tasks near the memory itself—no need for expensive round trips. Imagine running 10 servers’ worth of AI work on just one. That’s the kind of leap XCENA aims for.

Founded by veterans from Samsung and SK Hynix—the giants behind the world’s memory chips—XCENA is betting that memory is the next frontier. The company just raised $135 million at a $570 million valuation. Mass production is planned for late 2026, with revenue expected in 2027.

XCENA’s tech targets inference workloads where AI models juggle huge amounts of data outside of heavy matrix math. These chores include preprocessing, caching, and managing conversation context. By handling them inside memory, XCENA’s chip slashes power use and cost.

Tensormesh’s KV Caching Slashes AI Inference Costs

While XCENA rethinks hardware, Tensormesh attacks the software side. AI inference—running trained models to produce answers—is expensive because GPUs repeat the same work over and over. Every prompt often triggers a full recomputation of the AI’s entire context, wasting cycles.

Tensormesh’s secret weapon is “key-value caching,” or KV caching. It stores the intermediate data AI models generate while processing prompts. Instead of redoing calculations, the system reuses cached results instantly. This reduces latency and GPU costs by up to 10 times.

The company launched Tensormesh Inference, a SaaS platform that applies KV caching at scale. It offers real-time dashboards showing cost savings and cache hit rates. Some customers reach more than 70% cache hits, meaning most requests pull from memory, not compute. That means big bucks saved.

Backing this vision are Nvidia, AMD, and CoreWeave—three titans of AI hardware and cloud infrastructure. Their $20 million investment pushes Tensormesh’s total funding to $24.5 million. Tensormesh’s CEO calls KV caching a new kind of AI data that transforms inference economics.

The platform is flexible. Users can tap a serverless API compatible with OpenAI standards or opt for dedicated deployments with custom SLAs. Tensormesh also commits to open source, contributing to LMCache, the caching project it co-created.

Exa Labs Powers AI Search with Web-Scale Memory Efficiency

Another player in this memory-driven AI wave is Exa Labs. It’s building a search engine optimized for AI agents, not humans. Traditional search engines struggle to serve the massive, precise, and fresh data AI models need. Exa’s platform crawls over 500 billion URLs and uses token-efficient methods to speed up queries.

Exa just raised $250 million in Series C funding, soaring to a $2.2 billion valuation in under a year. It powers over 5,000 companies and 400,000 developers with low-latency, structured search results tailored for AI. Its tech reduces the tokens needed per search by up to 20 times, cutting costs and speeding responses.

This funding will expand Exa’s infrastructure, accelerate model training, and grow its team with top hires from Google, Meta, and other tech leaders. Exa aims to dominate the AI-native search layer, which will be critical as AI agents conduct vastly more searches than humans ever could.

The Memory-Centric AI Future is Here

These breakthroughs show AI hardware and software are shifting focus from raw compute power to clever memory use. XCENA’s chip brings processing power inside memory modules. Tensormesh’s KV caching eliminates repeated computations. Exa Labs refines search for AI’s massive scale.

The impact is clear. Hyperscalers spending billions on AI infrastructure crave every efficiency. Memory innovations can save hundreds of millions of dollars and unlock faster, smarter AI products. The AI revolution is no longer about just building bigger chips. It’s about rethinking how data flows and lives inside the system.

Keep your eyes peeled. Memory-centric architectures will shape AI’s next leap. The startups and giants backing these ideas are rewriting the rules of AI economics. The AI systems of tomorrow will run smarter, faster, and cheaper because of these bold moves today.