Jon Peddie Research summarized Intel's shift bluntly: the company has moved away from high-end discrete gaming GPUs and now prioritizes integrated graphics and high-margin data center solutions .
The net result for PC gamers: the Arc B580, launched in December 2024, will likely remain Intel's last new consumer discrete gaming GPU until at least late 2027 .
Intel hasn't abandoned GPU silicon — it has repurposed every design that was once destined for gaming toward professional and AI workloads, where margins are higher and the competitive dynamics with Nvidia play out differently.
The clearest example is the Arc Pro B70, which launched on March 25, 2026. It uses the same BMG-G31 die that was originally rumored to power the canceled Arc B770 gaming card. But instead of a 16 GB gaming card, it shipped as a 32 GB workstation GPU priced at $949 .
Key specifications:
Intel positions the Arc Pro B70 directly against Nvidia's RTX Pro 4000 ($1,800) — offering twice the VRAM (32 GB vs. 16 GB) at roughly half the price . The company's own marketing materials, datasheets, and product pages emphasize AI inference, multi-GPU scalability for large language models, content creation, and engineering workloads
. Gaming is never mentioned.
At Computex 2026, Sparkle demonstrated a single-slot blower-cooled version of the Arc Pro B70, enabling a single workstation to house 8 cards with a combined 256 GB of VRAM — enough to run 200-billion-parameter large language models locally . The card's total graphics power drops to 160 W in this configuration, down from 230 W in the standard dual-slot version, trading some sustained performance for dense multi-GPU deployment
.
Intel also used Arc Pro B70 and B65 GPUs in its MLPerf Inference v6.0 submission, demonstrating a 4-GPU system with 128 GB of total VRAM running 120-billion-parameter models at high concurrency . The Arc Pro B70 delivered up to 1.8x higher inference performance than the previous-generation Arc Pro B60
.
Intel's most ambitious AI GPU revealed at Computex 2026 was Crescent Island — a data center inference accelerator built on the Xe3P architecture, the same core design as the canceled Celestial gaming GPUs .
Where Crescent Island departs from industry convention is its memory. Instead of using expensive, supply-constrained High Bandwidth Memory (HBM) — which powers Nvidia's Blackwell GPUs — Crescent Island uses LPDDR5X memory in capacities up to 480 GB per card . This directly addresses a major pain point: the HBM supply chain is stretched thin, and enterprises that need more inference capacity often cannot get it because Nvidia's GPUs are both costly and scarce
.
Other key design choices:
Intel has not published raw throughput benchmarks for Crescent Island, making direct performance comparisons with Nvidia's accelerators impossible at this stage . The product's appeal is its memory capacity and cost structure — not peak compute, which Intel isn't claiming to lead on.
Intel's GPU strategy pivot is happening under direct competitive pressure from Nvidia. At the same Computex 2026 show, Nvidia CEO Jensen Huang unveiled the RTX Spark Superchip — a system-on-chip that combines 20 Arm CPU cores with a Blackwell-class GPU and up to 128 GB of unified LPDDR5X memory, manufactured on TSMC's 3 nm process .
RTX Spark is Nvidia's first end-to-end platform play for Windows laptops, co-developed with Microsoft and MediaTek, and heading to six major OEMs this fall . It challenges Intel across CPU, GPU, and AI acceleration in a single package — and represents the first time Nvidia's superchip architecture has appeared in consumer laptops
.
The market response was swift: Intel shares fell roughly 3.7% on the announcement, while AMD dropped 5.5% and Qualcomm tumbled 8.9% . Analysts described Intel as facing the "most acute long-term threat" from RTX Spark, squeezed between Apple Silicon on the premium end and Nvidia on the AI-performance end
.
Intel's public response was diplomatically measured but revealing. Nish Neelalojanan, senior director of product management for Intel's Client Computing Group, told Tom's Hardware at Computex 2026 that the company has "a healthy dose of paranoia" about RTX Spark . He emphasized Intel's confidence in x86 compatibility and suggested that Nvidia's Arm-based platform would face the same kernel-mode driver emulation and DRM challenges that have affected Windows on Arm gaming
.
Intel is also betting on price segmentation: its Wildcat Lake processors will start at $599, potentially undercutting Nvidia's premium-priced RTX Spark systems .
Intel's broader GPU strategy now revolves around three pillars beyond discrete gaming: integrated graphics in client CPUs, edge AI platforms, and robotics.
Intel has not officially declared an end to its gaming GPU ambitions. At Computex 2026, a senior director told PC Guide that discrete GPUs remain "super important" to the company's PC business . But the product record tells a different story: every new Arc design revealed in 2026 — the Pro B70, the Pro B65, Crescent Island, and the integrated Xe3P graphics in Panther Lake — is built for AI inference, workstations, or data centers.
The silicon that was supposed to power Intel's next gaming generations — BMG-G31 for the B770, Xe3P for Celestial — has been redirected to professional products instead. And the one gaming card Intel did ship recently, the Arc B580, now looks like an endpoint rather than a stepping stone.
For the discrete GPU gaming market, this leaves Nvidia and AMD as the only two competitors for the foreseeable future. Intel's retreat from gaming GPUs isn't a technology failure — Battlemage silicon works and is shipping in the Pro B70 — but a business decision to pursue higher-margin AI and professional workloads at a time when Nvidia is attacking Intel's core PC business from a completely new angle.
Comments
0 comments