Guest: Adi Gelvan
Company: Speedata
Show: The Agentic Enterprise
Topic: AI Infrastructure

Enterprise AI has a problem that no one in the boardroom wants to talk about. The models are ready. The investment is committed. But the infrastructure underneath—the systems responsible for moving, cleaning, and processing the data those models need—is cracking under pressure. Costs are ballooning, energy consumption is unsustainable, and the hardware powering analytics was never designed for this moment.

Adi Gelvan, CEO of Speedata, joined TFiR to explain what’s actually holding enterprise AI back, and why the answer lies not in software, but in silicon.

A New Class of Processor for a New Era

For four decades, the CPU was the default engine for computing. Then the GPU emerged, first for graphics, and later as the backbone of AI model training thanks to its strength in matrix multiplication. But neither was designed with data analytics in mind.

“There was never built a silicon intended to do data processing,” Gelvan said. “As AI scales, data becomes the bottleneck, and an APU—which is silicon purpose-built to run analytics—is the solution for this period.”

Speedata’s APU executes SQL queries directly on chip, a fundamental departure from the general-purpose architectures organizations currently rely on. The result is a processor purpose-built to handle the exact operations—batch ETL, data lake processing, and AI pipeline preparation—that bring CPU and GPU clusters to their knees.

The Numbers That Change the Conversation

Abstract architectural claims are easy to make. The customer results are harder to argue with.

One of Speedata’s largest customers, a major ad tech company, ran a direct comparison between a 38-server production cluster and three servers equipped with six APUs. The outcome: 50% faster processing with less than 10% of the original hardware. That customer is now planning to replace a 1,200-node cluster with fewer than 100 servers.

“From power consumption, footprint, and performance—it was a huge cost savings,” Gelvan noted. “You can actually put more and more nodes and servers in the data center without increasing the data center’s footprint.”

Where the APU Fits: Three Critical Workloads

Gelvan outlined three workloads where APUs are already proving transformative.

The first is batch ETL—the foundational work of moving data from databases into data lakes and warehouses. This is the classic analytics use case, and the one Speedata was originally designed to serve.

The second is AI data pre-processing. Before any model can be trained, raw data must be cleaned, transformed, and structured. Gelvan estimates that 70 to 80% of the time spent before model training is consumed by this work. APUs accelerate the SQL-heavy operations at the core of that process, making model training faster without touching the model itself.

The third workload—one that Gelvan describes as defining the next phase of enterprise AI—is retrieval-augmented generation (RAG). Unlike consumer chatbots that hallucinate and produce unpredictable outputs, enterprise AI requires access to structured, proprietary data. LLMs achieve this through RAG, spawning SQL queries against internal databases. Running those queries fast and at scale demands purpose-built hardware. “That’s where Speedata comes into play,” Gelvan said. “Our APUs enable companies to use LLMs while accessing their structured data—and we think the next stage of the AI revolution is enterprise adoption, which will need APUs.”

Sovereign Cloud and the Global Data Sovereignty Shift

Beyond performance and cost, there is a geopolitical dimension to Speedata’s growth story. The company recently announced a partnership with Nilo Cloud, a Dutch managed cloud provider, to bring APU-powered data processing to European sovereign cloud infrastructure.

The driver is clear: as countries and regions increasingly demand that data remain within their borders, reliance on hyperscale public cloud providers like Snowflake or Databricks becomes politically and operationally complicated. Gelvan sees this trend accelerating globally.

“Many countries are investing in sovereign clouds where data can reside within the country or continent. We believe that phenomenon is going to spread across the world—we’re already seeing it in Asia Pacific.”

Hardware Is Hard—and That’s the Moat

Speedata’s R&D is based in Israel, and the company manufactures its APU chips through TSMC before assembling the final PCIe cards for customer deployment. Gelvan is candid about the complexity involved—chip design cycles are long, supply chains are fragile, and the current surge in memory costs has made custom silicon even more expensive to build.

But Speedata’s architecture works in its favor. The APU is designed to operate with significantly less memory than comparable CPU or GPU-based systems, which Gelvan believes will become a structural cost advantage as memory prices remain elevated.

The Sustainability Case

Energy consumption is increasingly a boardroom-level concern for any organization scaling AI. Gelvan frames this not just as an environmental issue but as an adoption blocker.

“GPUs are not energy efficient, and if you use CPUs, you need a lot of servers—so they’re also not energy efficient. The way we’ve designed our technology is to be very cost-efficient on memory, footprint, and energy. We think that will help sustain this market and enable AI adoption, which is becoming a really big problem.”