Intel

Overview

Intel occupies a position no other company in this market shares: it designs processors, and it owns leading-edge fabrication capacity. Every other name tracked on this site depends on TSMC or Samsung to build its silicon. That dual role makes Intel simultaneously a competitor to the accelerator vendors and a potential supplier to them.

For most of the current AI cycle Intel has been the conspicuous laggard. Its Gaudi accelerator line, acquired with Habana Labs, never achieved the deployment scale of Nvidia or AMD parts, and the company's AI messaging has changed direction more than once. Under chief executive Lip-Bu Tan, appointed in 2025, the strategy has narrowed to something more defensible: rather than contest the frontier training market, Intel is targeting cost- and power-constrained inference, where the buyer cares about performance per dollar and per watt in an ordinary air-cooled rack.

The financial picture has improved markedly. Second-quarter 2026 revenue was $16.1 billion, up 25% year over year — the company's fastest growth in more than fifteen years. The Data Center and AI segment produced $6.3 billion, up 59%, and Intel Foundry $5.8 billion, up 31%.

Chip and product families

Role in inference: the memory bet

Crescent Island is worth examining because its central design decision runs against the industry consensus. Nearly every serious inference accelerator uses HBM, on the reasoning that decode is memory-bandwidth-bound and HBM provides the most bandwidth available. Intel has chosen LPDDR5X instead, and fitted 160GB of it.

That trade is coherent for a specific and fairly large class of buyer. LPDDR5X offers much less bandwidth than HBM but far more capacity per dollar, draws less power, and does not require liquid cooling or a place in the HBM allocation queue. A great deal of real-world inference is not frontier-scale: it is mid-sized models, served at moderate concurrency, by enterprises with conventional air-cooled data centres and finite budgets. For that buyer, a card that holds a large model in memory, fits an existing rack and can actually be purchased may beat a theoretically faster part that is expensive, power-hungry and back-ordered.

Where it will not compete is frontier serving. If the workload is a very large model at high concurrency with aggressive latency targets, bandwidth wins and Crescent Island is the wrong tool. Intel appears to understand this — the positioning is explicitly about performance per dollar and per watt, not peak throughput.

Strengths and key questions

Leadership and the SambaNova relationship

Lip-Bu Tan became chief executive in 2025, arriving from a long career in semiconductor investing through Walden International and a period leading Cadence Design Systems. His stated framing is that Intel should capture growth across its CPU franchise, ASICs, advanced packaging and foundry network — a portfolio position rather than a bet on beating Nvidia at accelerators.

Intel's relationship with SambaNova is unusual enough to note. Tan chairs SambaNova, a position predating his Intel appointment, and Intel holds a minority stake of roughly 9% following an additional investment in early 2026, alongside a strategic collaboration announced that February. The two companies also signed a non-binding term sheet for Intel to acquire SambaNova at a valuation of roughly $1.6 billion. That deal lapsed, and SambaNova subsequently raised $1 billion at an $11 billion valuation in July 2026 — roughly seven times the price Intel had contemplated. Coverage that still describes an Intel acquisition of SambaNova as pending is out of date.

Financials

Intel does not separately disclose Gaudi or accelerator revenue within the Data Center and AI segment, which also contains Xeon. Accelerator-specific revenue is therefore undisclosed.

What to watch

• Crescent Island sampling feedback. Independent measurement of how a 160GB LPDDR5X part actually performs on realistic mid-size inference workloads.
• 18A yield and external customers. Whether Intel Foundry wins accelerator designs from companies that currently use TSMC.
• Roadmap continuity. Whether Crescent Island gets a stated successor, which would signal commitment rather than another one-off.
• 14A in 2028. The commitment Tan has made publicly, and the one that determines whether Intel remains at the leading edge.
• Software investment. Whether Intel funds an inference software effort proportionate to the hardware.

Sources

1. Intel Investor Relations — Second-quarter 2026 financial results. Revenue, segment figures and guidance.
2. Tom's Hardware — Crescent Island, an inference-only GPU with Xe3P and 160GB of memory. Architecture and memory configuration.
3. HPCwire — Intel unveils new data centre GPU for inference. Original announcement and positioning.
4. Intel — Gaudi AI accelerators. Product documentation.
5. Fierce Network — Data centre and AI unit drives highest revenue growth in 15 years. Independent reporting on Q2 2026.
6. TechCrunch — SambaNova raises $1B at an $11B valuation (8 July 2026). Context on the lapsed Intel acquisition.

Financial figures are from Intel's own reporting. Crescent Island specifications are as announced; the product had not reached general availability at the time of review. See the editorial methodology.

Related pages