a16zFormer Intel CEO: Why This is the Best Time to Build Hardware
At a glance
WHAT IT’S REALLY ABOUT
AI revives hardware, but power, memory, and manufacturing now bottleneck
- Pat Gelsinger describes a hardware “renaissance” where AI accelerates chip design, but fabrication, packaging, and rack-scale integration now dominate time-to-value.
- He argues that today’s explosion of AI inference accelerators will consolidate as workloads evolve, capital requirements bite, and major platforms pick a few hardware/software ecosystems to scale.
- He calls HBM the best available but fundamentally flawed, and predicts memory innovation—new materials and compute-memory co-design—will become unavoidable because AI is a memory-centric workload.
- He expects 3D stacking to advance but remain constrained by yield, thermals, and power delivery realities, making moderate stack heights the practical sweet spot.
- He claims energy capacity is becoming the binding constraint on AI expansion, pushing innovation in generation (including nuclear), high-voltage DC data centers, power electronics, cooling, and eventually optical interconnect plus more flow/circuit-like networking.
IDEAS WORTH REMEMBERING
5 ideasAI is shrinking chip-design time; fabrication and packaging are now the real bottlenecks.
Gelsinger argues that AI makes the *design* step dramatically faster (months), but the system still can’t ship value until manufacturing, advanced packaging, and rack integration happen—often stretching to ~9 months or more. The practical frontier is reducing fab cycle time, mask costs, and packaging lead times so hardware can track rapidly changing AI workloads.
The “100 AI chips” moment is temporary; the market will converge to a few platforms.
He expects the current proliferation of AI accelerator startups to narrow because (1) over-specializing for transient workload phases (prefill/decode/mid-fill, etc.) becomes operationally unsustainable, (2) capital and scale requirements eliminate most entrants, and (3) major ecosystem players will select and software-enable a small number of “winners,” hiding heterogeneity behind abstraction layers.
AI turns memory into the central constraint—and may finally trigger real memory innovation.
Calling HBM “hideous” but necessary, Gelsinger highlights poor bit density, thermal issues (DRAM hates heat), bandwidth constraints (“shoreline bandwidth”), and power. He believes AI’s memory-centric nature and the new profitability of memory suppliers will finally fund new materials and architectures (e.g., ferroelectrics, non-capacitive, stackable high-density approaches).
3D packaging will grow, but manufacturing yield will cap stack height at modest levels.
He is skeptical that very tall stacks (e.g., 16–32 layers) will be practical because yield must improve superlinearly/exponentially as stack height increases; cracked dies can’t be rescued by redundancy. He predicts a “sweet spot” of modest logic/memory stacking (often 2–4 high for memory) plus added layers for power delivery, redistribution, and eventually integrated optics.
Optical is the future for I/O and networking—just not for the core compute-memory datapath.
Gelsinger supports optics broadly for I/O and cluster interconnect but dislikes heavy optical/electrical conversions inside the core compute-memory complex due to power losses: communication energy per bit is orders of magnitude worse than compute energy per operation. He predicts in-package optics and more optical switching (more circuit/flow-oriented) around 2028–2029, driven by predictable, large AI flows and the limits/cost of copper at scale.
WORDS WORTH SAVING
5 quotesIn a AI digital age, energy capacity equals economic capacity. Why build the new data center and buy the million GPUs if I can't power them? You're gonna see more and more defaults happening on many of those data center projects because the energy won't be there.
— Pat Gelsinger
W- whenever you have the technology to make something easy, that means the bottleneck moves somewhere else.
— Guido Appenzeller
Nothing's a chip anymore, it's a rack. It took me three months to design it, but it's nine months until I can actually start to use it.
— Pat Gelsinger
HBM is a hideous memory. It's just the best one that we got.
— Pat Gelsinger
I declared the death of copper about 25 years ago. Eventually I'll be right.
— Pat Gelsinger
High quality AI-generated summary created from speaker-labeled transcript.
