TechnicalDiscussion

What will computer chips look like in 2045?

Dwarkesh Podcast1h 9m

Intel faces unprecedented AI-driven demand requiring 40-50% more computing capacity than maximum capability. The company must balance aggressive technology innovation with predictable manufacturing schedules while competing with TSMC, constrained by upstream tool suppliers like ASML who produce fewer than 100 EUV machines annually.

Summary

The transcript documents Intel's strategy for advancing semiconductor manufacturing in response to explosive AI demand. Chris Oth, who leads process development, explains that Intel's 14A process—launching in 2028—has been in development for three-plus years with approximately 2,000 manufacturing steps requiring intricate fine-tuning. The company learned critical lessons from the 10-nanometer node, which experienced four-year delays due to excessive innovation bundled into a single generation. Intel now uses 'phase gate' maturity metrics to decide which technologies are ready for production, prioritizing schedule reliability over maximum innovation.

The discussion covers transistor evolution from planar designs to FinFET (2011) to gate-all-around technologies (2022), showing how transistor innovations take decades to mature despite lab demonstrations occurring earlier. Gate-all-around required solving material science challenges, including high-K dielectrics, new etching processes, and thermal management. Intel's R&D shifted from 1980s-90s simple scaling to a 'golden age of innovation' requiring simultaneous advancement across multiple domains: materials increased 10x, atomic-layer deposition enables atom-by-atom control, and new architectural approaches like backside power delivery emerge.

For 14A specifically, Intel adopts high-NA EUV lithography while TSMC deems it uneconomical. This divergence stems from Intel's backside power architecture, which reduces front-side pitch requirements, making high-NA's 1.7x feature reduction beneficial. Without backside power, TSMC must use double-patterning, negating high-NA's advantages.

Intel's transition to foundry services required building formal Process Design Kits (PDKs), a new capability after decades as an integrated device manufacturer. 18A PDK quality issues created customer friction; 14A addresses this with maturity metrics and higher quality from launch.

Kevin Fisher describes 'Hot Topics' meetings where engineers propose experiments and integration leads make resource allocation decisions under extreme time pressure. One example: epi nodules during source-drain deposition caused transistor shorts; the team pursued multiple simultaneous fixes rather than waiting for root cause analysis, using tiered 'hot box' factory prioritization to accelerate critical experiments.

The research lab explores 2D materials, CFET (stacked complementary transistors), and ruthenium interconnects. CFET stacking increases aspect ratios and interconnect complexity, requiring simplified test vehicles before full implementation. Ruthenium addresses copper's resistivity issues at sub-20nm pitches by eliminating liner layers. Tyler Osborne explains hybrid bonding technology enables sub-micron pitch connections between stacked chips, with literature demonstrating 0.45-micron pitch reliability and potential future sub-0.5-micron scaling.

Pushkar Arande, Intel's CTO, identifies the critical bottleneck: ASML produces fewer than 100 EUV tools yearly with no plans to significantly increase production this decade. All equipment faces supply shortages; once processes lock into specific tools, switching becomes impossible. Even with unlimited CapEx, Intel estimates only 2x maximum wafer production scaling possible—far below the 3x year-over-year AI compute demand trajectory. TSMC produces 3x more advanced node wafers than Intel, leveraging economies of scale where 80% of fab costs are fixed; therefore, each utilization percentage gain multiplies revenue disproportionately.

Arande notes that catching up to TSMC requires scale gradually; current performance differences are 'a few percent,' which matters less than predictable execution, high yields, and supporting rapid ramps. He disputes semi-analysis claims that 14A transistor density matches TSMC's N2, emphasizing that different transistor types (low-leakage, high-performance, SRAM variants) and analog/RF circuits make competitiveness multidimensional.

A paradoxical insight emerges: AI demand could generate $100 billion annual revenue from a few billion dollars of Intel CapEx, yet supply chain constraints prevent scaling. Industry consolidation reduced foundries from 20 to 3, driven by extreme economies of scale. The value chain is shifting from horizontal (separate foundries, design, equipment companies) toward 'pseudo vertical integration' where hyperscalers design custom chips and guarantee entire fab capacity.

Arande speculates on radical rethinking: questioning clean room necessity, rethinking transistor architecture, and reconsidering manufacturing processes locked in 30-40 years ago. However, semiconductors' capital intensity and high failure costs prevent incumbent companies from experimenting; someone must 'start with a clean slate.'

Intel's TeraFab partnership with Elon Musk addresses this bottleneck partially—new fabs could guarantee tool supply but face 18-month tool delivery and multi-year construction timelines extending into 2030. The conclusion emphasizes that had Intel known current demand levels a year ago, it would have invested more in tool commitments; recognizing this now, the company should commit billions to secure tools before Taiwan-based suppliers become the limiting factor.

About this episode

<p>Demand for AI compute has grown 3x a year for the last decade, but according to Intel’s CTO, there are only enough chipmaking tools in the world to roughly double leading-edge production.</p><p>I wanted to understand what innovations are coming out of the fabs, and what they mean for the future of computers, so I spent two days at Intel talking to the people working on them – from the head of process development to the research lab, the packaging group, and the CTO.</p><p>A lot of what’s coming next – gate-all-around transistors, backside power, CFET, hybrid bonding – is hard to follow without seeing it, so some of this video is the animated explainer I wish I had while I was prepping.</p><p>We haven’t done this format before, so let me know what you think!</p><p>I also made flashcards for the concepts I had to learn to prep for this episode <a href="https://flashcards.dwarkesh.com/" target="_blank">here</a>.</p><p>Watch on <a href="https://youtu.be/frYSOKvNiiE" target="_blank">YouTube</a>, read the <a href="https://www.dwarkesh.com/p/intel-tour" target="_blank">transcript</a>.</p><p>Sponsors</p><p><a href="https://workos.com" target="_blank">WorkOS</a> helps you do business with enterprises. It gives you access to all the features most commonly required by enterprises: SSO, SCIM, RBAC, audit logs, and more. It’s easy to plug these features in — it only takes a few lines of code — and WorkOS is releasing new functionality all the time, including things like auth for agents. If you want to join the 3,000+ companies that rely on WorkOS, go to<a href="https://workos.com" target="_blank"> </a><a href="http://workos.com" target="_blank">workos.com</a></p><p><a href="https://janestreet.com/dwarkesh" target="_blank">Jane Street</a> built a powerful tool called magic-trace to help analyze and debug their trading systems. It only exists because one engineer noticed a feature on Intel chips called Processor Trace and thought it’d be valuable to build on top of. Now, traders and developers across Jane Street use magic-trace every single day! If you want to be trusted to pursue the ideas you think matter most, Jane Street is hiring. Go to<a href="https://janestreet.com/dwarkesh" target="_blank"> janestreet.com/dwarkesh</a></p><p><a href="https://antithesis.com/dwarkesh" target="_blank">Antithesis</a>’ testing platform lets you perfectly reproduce every single bug you find. That’s because it runs on a fully deterministic computer, which strips out all the unpredictability that can make certain bugs flaky and hard to pin down. To build it, Antithesis found every source of non-determinism in Intel’s x86 architecture and either sealed it off or replaced it with a controlled simulation. Learn more at<a href="https://antithesis.com/dwarkesh" target="_blank"> antithesis.com/dwarkesh</a></p><p>Timestamps</p><p>(00:00:00) – Intro</p><p>(00:00:47) – How Intel brings up a new process node</p><p>(00:03:22) – What Intel learned from 10nm</p><p>(00:06:28) – From planar transistors to gate-all-around</p><p>(00:13:13) – Why High-NA EUV pays off for Intel but not TSMC</p><p>(00:19:20) – Could chips have evolved completely differently?</p><p>(00:23:33) – Inside Intel’s “Hot Topics” meeting</p><p>(00:29:30) – Replacing copper wires with ruthenium</p><p>(00:32:41) – Why stacked transistors take a decade to ship</p><p>(00:39:45) – Hybrid bonding and the chip-to-chip bottleneck</p><p>(00:48:00) – There aren’t enough tools to 10x chip production</p><p>(00:53:14) – Can Intel catch up to TSMC?</p><p>(01:01:02) – Terafab, and rethinking the fab from scratch</p> <br /><br />This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit <a href="https://www.dwarkesh.com?utm_medium=podcast&#38;utm_campaign=CTA_1">www.dwarkesh.com</a>

Key Insights

  • AI is driving 40-50% demand surge above Intel's maximum manufacturing capacity, an unprecedented shock to the semiconductor industry.
  • Intel learned from its 10-nanometer node delays (2016 target to 2018 production) that bundling too much innovation into one generation undermines schedule predictability, leading to a new strategy of limited innovation per node.
  • Gate-all-around transistor technology was demonstrated in 1988 but didn't reach volume production until 2022, showing that lab demonstration and manufacturing maturity are separated by decades of material science and process development work.
  • Intel's backside power architecture fundamentally changes lithography economics: high-NA EUV becomes cost-effective for Intel while TSMC finds it uneconomical because high-NA's benefits depend on reduced front-side pitch requirements.
  • ASML produces fewer than 100 EUV machines annually with no plans to significantly increase production this decade, making tool supply the hard constraint limiting global semiconductor output regardless of fab investment.
  • Even with unlimited capital expenditure, Intel estimates only 2x scaling of advanced-node wafer production is physically possible, far below the 3x year-over-year AI compute demand growth.
  • Semiconductor foundries operate with 80% fixed costs, making each percentage increase in utilization or yield multiplicatively valuable, which drives extreme economies of scale and industry consolidation to just three foundries.
  • Intel's transition to foundry services required building formal PDKs for external customers, a new capability after decades as an integrated device manufacturer, revealing gaps in process documentation and customer support.
  • The value chain is shifting from horizontal (separate foundries, chip designers, equipment makers) toward 'pseudo vertical integration' where hyperscalers like OpenAI and Anthropic design custom chips and guarantee entire fab capacity.
  • Ruthenium interconnect material becomes advantageous below 20-nanometer pitch because it eliminates copper's required liner layer, trading resistivity for simpler deposition without architectural modification.
  • Hybrid bonding technology has demonstrated 0.45-micron pitch reliability in published research with potential for sub-0.5-micron scaling, representing another order-of-magnitude density increase beyond current 25-micrometer pitch production.
  • The semiconductor industry's extreme risk-aversion prevents radical experimentation with manufacturing processes despite potentially lower costs, because a single day of lost productivity costs millions and incumbent companies cannot risk unknown outcomes.

Topics

AI demand surge and manufacturing capacity constraintsTransistor design evolution and material innovationProcess technology nodes and development timelinesEUV lithography and tool supplier bottlenecksFoundry business model and PDK requirementsManufacturing yield improvement methodologiesHybrid bonding and advanced packagingSupply chain economics and consolidationGate-all-around and CFET transistor architecturesBackside power delivery systemsPhase gate maturity metrics for technology adoptionRisk-aversion in semiconductor manufacturing

Transcript

There's never been as big a shock to the semiconductor industry as the current surge in demand from AI. AI is driving incredible computing demand. Easily 40 to 50 percent above our maximum capacity, which I've never seen before. Progress in AI has been driven by exponentially increasing compute. To sustain that progress, foundries like Intel will need to manufacture far more leading-edge compute, both by producing more chips and by improving what each chip can do. But the production of these chips is bottlenecked by tools that Intel buys from their suppliers, most famously, ASML's EUV machines. If they don't have the equipment, they can't make the chips here. I spent two days touring Intel's fabs to prepare for…

Full transcript available for MurmurCast members

Sign Up to Access

More from Dwarkesh Podcast

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.