Intel Xeon 6+ Clearwater Forest: 288 cores on 18A and the agentic AI CPU bet

What Clearwater Forest actually ships

Xeon 6+ is available now across four SKUs, from the 144-core Xeon 6960E+ to the flagship 6990E+ with 288 cores. Dual-socket configurations push the ceiling to 576 cores per server. Every model supports 12-channel DDR5-8000 memory, 96 PCIe Gen 5 lanes, 64 CXL 2.0 lanes, and TDPs between 300W and 450W. Platform compatibility carries over from existing LGA 7529 Xeon 69xxE/P infrastructure, which matters for adoption: Dell, HPE, Gigabyte, Asus, and Foxconn are among the OEM partners already listing systems.

Architecturally, this is Intel's most complex chiplet design to date. Twelve compute tiles on 18A each carry 24 Darkmont E-cores; they stack atop three active base tiles on Intel 3 and two I/O tiles on Intel 7, connected via Foveros Direct 3D face-to-face bonding and EMIB 2.5D interconnect. The result is 576 MB L3 alongside 288 MB L2 — roughly five times the prior generation's last-level cache. Intel SGX and TDX confidential-computing extensions are baked into silicon for multi-tenant and regulated workloads.

The 18A node itself is a process milestone, not marketing filler. Clearwater Forest is the first shipping data-center product on a node that combines RibbonFET gate-all-around transistors with PowerVia backside power delivery. For Intel's foundry turnaround story — and for Lip-Bu Tan's credibility twelve weeks into the CEO role — proving 18A in volume silicon matters as much as any benchmark slide.

Why Intel is pitching CPUs for agentic AI

Training-era clusters optimized for a familiar ratio: roughly one CPU per four GPUs, with the GPU doing nearly all meaningful compute. Agentic workloads invert parts of that math. Running tool-calling agents, retrieval pipelines, sandboxed code execution, and concurrent small-model inference stresses orchestration, memory bandwidth, and core count more than raw tensor throughput. Creative Strategies analyst Ben Bajarin, cited in Intel's keynote materials, frames the shift as moving toward roughly one CPU per GPU or less in agentic deployments.

Intel's headline density claim illustrates the pitch: a single liquid-cooled rack can deliver 36,864 cores in 32U of compute space at roughly 100 kW — marketed as the highest agent-hosting density available. Whether that figure survives independent benchmarking is an open question; what is not debatable is the strategic direction. Microsoft's Scout agent on open-source OpenClaw and enterprise control-plane investments show that software vendors are shipping always-on agents. Someone has to run the orchestration layer, memory management, and I/O fan-out those agents generate. Intel wants that layer on Xeon, not as an afterthought attached to a GPU tray.

The COMPUTEX stage also showcased rack-scale partnerships that extend beyond the CPU die. Intel, SambaNova, and Foxconn demonstrated production-ready racks combining Xeon processors with SambaNova SN-50 reconfigurable dataflow units for inference acceleration. Separately, Vector Core Compute — an enterprise inference cloud backed by Vista Equity Partners and Cambium Capital — ran a live demo of fully disaggregated inference: Xeon 6 for orchestration, SambaNova RDUs for decode, and NVIDIA Blackwell GPUs for prefill. Together.ai was named as the first commercial customer. The message is heterogeneity: the winning rack is not GPU-only or CPU-only, but deliberately split across prefill, decode, and control planes.

Market context: selloff, capex anxiety, and the inference pivot

Timing is everything. Intel announced Xeon 6+ availability the same week Broadcom guided below AI revenue expectations, wiping hundreds of billions from semiconductor market caps and amplifying Asia's foreign-selling spillover. Investors are no longer willing to fund every training-cluster expansion on faith. Oracle's June 10 earnings report — previewed in our OCI and RPO stress-test analysis — will be read as a referendum on whether cloud providers can convert booked AI demand into recognized revenue.

Intel's agentic pivot offers a narrative escape hatch: if the marginal dollar of AI capex shifts from training frontier models to serving production agents, CPU vendors regain relevance. E-core density at 450W TDP is a credible answer for cloud-native and telecom workloads that never needed a GPU in the first place. Foxconn's plan to manufacture a CPU-dense rack variant without additional accelerators — aimed at cost-optimized inference and hybrid AI — signals that tier-1 contract manufacturers see a volume path that does not require HBM stacks.

Skepticism remains warranted. Intel has shipped “AI optimized” Xeon generations before without displacing NVIDIA in the workloads that drive headlines. Clearwater Forest does not compete on LLM prefill; it competes on everything around prefill — session management, retrieval, policy enforcement, sandboxed tool execution, and the long tail of small models enterprises actually deploy. That is a smaller TAM than training, but it is also where inference serving economics become operational rather than experimental.

Three scenarios for Xeon 6+ adoption through 2027

Scenario A — Inference-rack standard for enterprise agents (40–45% probability): Hyperscalers and neoclouds adopt CPU-dense racks for agent orchestration alongside GPU prefill pools. Intel regains data-center share in telco/5G and cloud-native segments where E-core density already won on power efficiency. 18A yields hold, foundry customers cite Clearwater as a reference design, and Lip-Bu Tan's turnaround narrative gains Wall Street patience through at least two earnings cycles.

Scenario B — Niche win, GPU centrality unchanged (35–40% probability): Xeon 6+ sells into refresh cycles and edge-adjacent workloads but fails to shift the AI capex conversation. NVIDIA Blackwell and custom ASICs capture agent prefill/decode; CPUs remain necessary but not strategic. Intel's 18A milestone helps the foundry business on paper while client and data-center revenue stay under pressure. Stock trades on cost cuts, not growth.

Scenario C — Capex freeze delays rollout (15–20% probability): A deeper AI revenue disappointment — think Oracle miss plus weak CPI — pushes enterprises to extend existing Xeon 5 clusters rather than adopt 6+. OEM inventory builds, Intel discounts SKUs, and the agentic-AI CPU thesis gets shelved until 2028. 18A execution continues, but volume ramps slower than the COMPUTEX keynote implied.

What to watch next

• Independent inference benchmarks on 6990E+ versus AMD EPYC Turin and prior-gen Xeon for agent orchestration workloads (not just SPECrate).
• Oracle and Broadcom earnings June 10–11 for signals on whether cloud AI capex is reallocating toward inference racks or freezing entirely.
• Foxconn rack shipments and SambaNova SN-50 attach rates in production deployments, not keynote demos.
• 18A yield commentary on Intel's Q2 call — client Series 3 momentum is already tied to the same node.
• WWDC and enterprise agent announcements June 8–12 for software demand that would justify CPU-dense hosting economics.

Clearwater Forest will not settle the bitcoin-versus-AI rotation debate or reverse a Friday cross-asset selloff. It does mark a concrete product answer to a question the market had been asking in abstract terms: if agents are the next platform shift, what silicon runs the control plane? Intel's bet is 288 E-cores, half a gigabyte of L3, and a rack that treats orchestration as first-class compute. After years of GPU headlines, that is either a timely pivot or a expensive hedge against a training bubble Intel never owned. COMPUTEX shipped the silicon; the June earnings season will decide whether anyone is still buying the story.

Sources: Intel Newsroom — COMPUTEX 2026 AI innovations (June 2, 2026); Data Center Dynamics — Xeon 6+ agentic AI launch (June 2026); TweakTown — Clearwater Forest SKU details (June 2026); Futurum Group — Xeon 6+ architecture analysis (June 2026).