Intel diverts resources to Jaguar Shores as AMD gains ground

Intel diverts resources from its Falcon Shores AI accelerator to focus on the Jaguar Shores rack-scale system. Interim co-CEO Michelle Johnston Holthaus confirmed this strategic pivot during the fourth-quarter earnings call, explaining that Falcon Shores will only serve as an internal test chip to help develop the hardware and software foundations for its successor. This decision follows Intel’s third consecutive quarterly loss, which saw a net loss of $1.6 billion and revenue drop to $12.7 billion. The company reported that its AI business, Habana, generated only $0.4 billion in revenue, falling far below its targets. Because the company struggled to achieve its $500 million revenue target for Gaudi chips due to a product transition from Gaudi 2 to Gaudi 3 and software ease of use issues, it must now focus on a broad AI data center solution. Holthaus made this decision as one of her first major moves after Pat Gelsinger’s departure in December. She arrived at this move after receiving feedback from customers and discussions with Intel’s AI accelerator chip team regarding where the company stood from a competitive and execution perspective. I view this move as a retreat from the single-chip competition where NVIDIA maintains a massive lead. While Intel targets a 2027-2028 launch for Jaguar Shores, AMD pursues an accelerated roadmap that releases new GPUs every year. One system integrator, Mark III Systems, chose AMD over Intel for its AI center of excellence program because of Intel’s inability to provide formidable competition to NVIDIA. Intel’s difficulty in capturing market share even with Gaudi 3 shows the massive gap between standalone silicon and a complete AI infrastructure.

Supply constraints and the shift toward rack-scale

NVIDIA maintains a significant lead through its Blackwell architecture and its control over the high-bandwidth memory (HBM) supply chain. SK Hynix allocates 70% of its capacity to NVIDIA, while Micron provides 60% of its supply to the same chipmaker, leaving Samsung to deal with a squeezed share. These constraints create a massive barrier for competitors attempting to build high-performance single-chip solutions. The industry has moved toward rack-scale systems like the NVIDIA Grace Blackwell GB200 NVL72, which integrates 36 Grace CPUs and 72 Blackwell GPUs over the NVLink chip-to-chip interconnect. Major partners including Dell Technologies, Amazon Web Services, Microsoft Azure, Google Cloud, and Oracle Cloud offer this multi-node, liquid-cooled system. NVIDIA CEO Jensen Huang noted the complicated engineering required to integrate these systems with OEM and cloud partners. AMD positions itself as the primary alternative, claiming its Instinct MI325X can outperform the NVIDIA H200. You probably already know that NVIDIA’s head start makes the road to parity long, but AMD’s Instinct MI300X already sees customer interest for testing. The industry’s shift toward systems like the AMD Helios, which uses 72 MI400 GPUs, makes the single-chip competition against NVIDIA’s massive clusters increasingly irrelevant.

The new battleground for AI infrastructure

The battle for data center dominance centers on how well a company integrates networking, memory, and system-level optimizations into a single rack. AMD uses its Instinct accelerators to challenge the status quo, while Intel attempts to build a cohesive platform with Jaguar Shores. This strategy attempts to compete with massive rack-scale players like the AWS Trn3 UltraServer and the Google TPU 8t pod. To challenge NVIDIA’s proprietary NVLink ecosystem, AMD and Intel jointly drive the UALoF open interconnect. The market for AI training chips remains heavily skewed toward established players, as companies struggle to invest time in training teams around new products like Gaudi.

Intel admits it lacks a meaningful presence in the cloud-based AI data center market, a reality that complicates its transition from single-chip products to integrated systems. The company’s previous plan for a 200 PFLOPS single-chip Falcon Shores target proved physically impossible due to constraints in power, area, and HBM capacity. This failure leaves NVIDIA and AMD in control of the most compute-intensive workloads. AMD continues to expand its data center presence through its Instinct and EPYC brands. Can Intel successfully leverage its 18A foundry services to become a primary supplier for companies like NVIDIA or AMD? AMD remains the strongest contender to capture market share as it expands its data center presence.