Oct 9, 2026, 10:30 AM ETFireside Chat, Interview, Conference Presentation

Former Intel CEO: Why This is the Best Time to Build Hardware

Energy capacity is now the defining constraint on economic capacity in the AI era; a lack of power prevents data centers from justifying the purchase of millions of GPUs.
Pat Gelsinger predicts a wave of defaults on data center projects specifically due to the inability to secure sufficient power infrastructure.
Historical energy grid stagnation in the US (flatlining capacity for 10–15 years while coal was retired) has created a critical supply gap that must be filled by 4% annual growth to support AI demand.
Key power innovations required include the adoption of 800-volt DC distribution, solid-state transformers, and the revival of nuclear power as a base-load source.
Current power simulation tools are insufficient, forcing engineers to guard-band power envelopes by nearly 40%, representing a massive efficiency loss.

While AI tools can compress chip design time to approximately three months, the physical realization of the chip remains a nine-month bottleneck due to silicon processing and packaging lead times.
The industry faces a disconnect where AI-optimized designs are rendered obsolete by the time they reach scale due to rapid shifts in model workloads and agent architectures.
Current HBM memory is described as "hideous" due to limitations in bit density, shoreline bandwidth, and thermal constraints, necessitating a return to innovation after 30 years of stagnation.
Gelsinger forecasts the first major memory innovation in three decades driven by new materials like ferroelectrics and stacked memory structures, rather than traditional DRAM or Flash.
Vertical stacking beyond three to four layers is viewed as a manufacturing nightmare due to exponential yield requirements; two or four layers are identified as the "sweet spot."
Optical I/O is predicted to become dominant in 2028–2029, replacing copper for scale-up environments as copper transmission becomes economically and physically prohibitive at longer distances.
Gelsinger remains skeptical of 16+ layer stacks and PIM (Processing-in-Memory) architectures, favoring high-density, high-bandwidth memory placed physically close to compute engines.

Despite the current existence of roughly 100 AI inference accelerator startups, Gelsinger predicts a rapid convergence to a few dominant vendors, citing historical precedents where no industry sustains 100 competing processor vendors.
Extreme hardware specialization for specific AI sub-tasks (e.g., pre-fill vs. mid-fill) is viewed as unsustainable as workloads moderate, migrate, and require high-precision HPC-like calculations again.
Large cloud consumers (e.g., OpenAI, NVIDIA, Anthropic) will likely consolidate the field by selecting a limited number of platforms, prioritizing deep hardware-software co-evolution over pure hardware novelty.
Software constraints are eroding due to AI agents that can write optimized kernels for new architectures overnight, potentially allowing for slightly more heterogeneity in the short term before capital forces consolidation.
The distinction between training and inference is expected to blur as continuous learning algorithms update model weights within inference environments, reducing the need for distinct specialized hardware.

Network architecture is shifting from packet-switched models to optical circuit-switched or optical switching (OCS) models to accommodate the predictable, large-flow nature of AI workloads.
Gelsinger anticipates the "death of copper" for I/O interfaces, declaring it a 25-year-old prediction that will finally materialize as optical supply chains mature.
The virtual machine (VM) abstraction is undergoing a rebirth specifically for AI agents, requiring new management layers focused on agent security, migration, and policy enforcement rather than just hardware virtualization.
Future virtualization must balance human policy control (guardrails/constitutions) with agent performance needs, creating a dual-hierarchy for managing agent swarms.
Innovation in cooling and thermals is becoming a primary engineering focus, with new materials (e.g., diamond) and liquid cooling techniques essential for managing high-density compute stacks.

Gelsinger states, "Whenever you have the technology to make something easy, that means the bottleneck moves somewhere else," identifying manufacturing lead times and power as the new barriers.
He asserts that the current "renaissance" in hardware engineering is comparable to the intellectual fervor of the Italian Renaissance, driven by the need to solve complex power, thermal, and memory challenges.
The speaker forecasts that the memory industry, previously plagued by volatile commoditization cycles, has become a top-tier sector with the market cap of major tech companies, incentivizing the capital needed for R&D.
Gelsinger expects the industry to move toward 3D-integrated platforms where power rails, optical I/O, and memory are co-designed within a single eight-layer construct.
He predicts that failure to secure energy capacity will become the primary filter for data center viability, overriding other design considerations in the near future.