Intel Killed Its Gaming GPUs. Here's What It's Building Instead

At Computex 2026, Intel told reporters that discrete GPUs remain a "super important" part of its PC business. But the company didn't show a single new gaming graphics card — and for good reason. Behind the scenes, Intel has systematically dismantled its consumer gaming GPU roadmap. The Arc B770 is canceled. The next-generation Celestial architecture was scrapped "a long time ago." The planned Druid follow-up in 2027 is unconfirmed. Every piece of new Arc silicon Intel has actually shipped or announced in 2026 is built for AI inference, professional workstations, or data centers.

The pivot is stark, and the evidence is now overwhelming. Here's exactly what happened to Intel's Arc roadmap — and what's replacing it.

The Gaming Roadmap: Canceled in Three Stages

Intel's consumer gaming GPU ambitions have unraveled in a cascade of leaks, no-shows, and partner confirmations throughout 2026. The last new gaming product was the mid-range Arc B580, which launched in December 2024 at an attractive $249 price point and was generally well-received . Everything above it has since been killed.

The cancellations broke into public view in three distinct waves:

• Arc B770 (Big Battlemage): The long-rumored high-end Battlemage card based on the larger BMG-G31 die was a no-show at CES 2026 in January, and again at Computex 2026 in June. Partners have confirmed its cancellation, and Intel itself offered no roadmap updates for it at either event . The chip was originally rumored to feature 32 Xe2 cores, 16 GB of GDDR6 memory, and a 300 W TDP — positioned to compete with Nvidia's RTX 4070 — but it never materialized . Multiple media outlets now describe the B770 as effectively canceled .
• Celestial (Xe3P): On April 24, 2026, reputable Intel leaker Jaykihn reported that the entire Xe3P "Celestial" gaming GPU lineup had been canceled, and had been for "a long time" . This was not a delay or a pause — multiple outlets independently confirmed the cancellation . Intel has offered no public denial.
• Druid (Xe4): The planned 2027–2028 successor to Celestial has no confirmed discrete gaming variant. Its status remains entirely unconfirmed, leaving no visibility on a future consumer GPU from Intel for years .

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 .

What Intel Is Building Instead

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.

Arc Pro B70: The Gaming GPU That Became a $949 AI Card

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:

• 32 Xe2 cores, 256 XMX AI engines
• 32 GB GDDR6 memory on a 256-bit bus (608 GB/s bandwidth)
• 367 INT8 TOPS of AI inference performance
• 22.94 TFLOPS of FP32 compute
• PCIe 5.0 x16 interface, up to four DisplayPort 2.1 outputs

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 .

Crescent Island: A Data Center GPU That Bypasses the HBM Bottleneck

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:

• 350 W air-cooled PCIe form factor — no liquid cooling required, fitting standard enterprise servers .
• Targets "agentic AI" inference workloads, not training, with data type support from FP4 through FP64 .
• Customer sampling begins in the second half of 2026, with limited shipments targeted by the end of 2026 .
• A successor code-named Jaguar Shores is planned for 2027 .

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.

The Shadow of Nvidia's RTX Spark

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 .

Edge AI, Robotics, and the Integrated Graphics Play

Intel's broader GPU strategy now revolves around three pillars beyond discrete gaming: integrated graphics in client CPUs, edge AI platforms, and robotics.

• Panther Lake — Intel's next-generation client CPU built on the 18A process — will feature integrated Xe3P graphics cores with AI acceleration in mainstream laptops. These are iGPUs, not discrete cards, but they carry forward the same architectural DNA as Celestial and Crescent Island .
• OpenVINO and Intel's edge AI inference platforms received significant promotion at Computex 2026, with the company positioning its GPU IP for robotics and industrial AI workloads .
• Jon Peddie Research's analysis in May 2026 confirmed the strategy: Intel now prioritizes integrated graphics and high-margin data center solutions over consumer gaming discrete GPUs .

What This Means for the GPU Market

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.