Gilder's 'Trillion Dollar Triangle' — The Six Stocks Behind the Wafer-Scale Computing Pitch

The Hook

George Gilder has been predicting technology revolutions for five decades. His latest pitch, updated now that one of his “Titans” has finally gone public, argues that the next one is already underway: wafer-scale computing. The claim is that entire silicon wafers, used as single interconnected “super-chips” instead of being diced into individual processors, will process data in minutes that conventional GPU-based AI systems take “hours or days” to handle, all while using “up to 90% less energy.”

The implications, as Gilder frames them, are staggering. “These massive AI data centers are now write-offs waiting to happen.” The hundreds of billions of dollars pouring into hyperscale data centers, the electricity shortages that have made utility stocks into AI plays, the NVIDIA GPU monopoly, all of it, according to this thesis, is about to be rendered obsolete by three companies converging on what insiders call the “Trillion Dollar Triangle.”

This is a re-air. StockGumshoe covered the same pitch earlier this year, before Cerebras Systems (CBRS) went public in May at $185 per share. The ad copy has been updated to reflect the IPO, but the valuation data sprinkled throughout remains notably stale. At $77 for the first year of Gilder’s Technology Report, the price of admission is modest, but the claims are anything but.

The Big Claim

The “Trillion Dollar Triangle” rests on a straightforward technological argument. Current AI computing relies on GPUs: thousands of individual chips wired together, with data constantly moving between chips, across interconnects, and through memory hierarchies. Each hop costs time and energy. Wafer-scale computing eliminates those hops by keeping everything on a single silicon wafer, with direct on-wafer connections that avoid the need for separate packaging and inter-chip wiring. Less wiring means less latency, less heat, and dramatically lower power consumption.

The three Titans, as Gilder describes them: Titan #1 pioneered the wafer-scale architecture and proved it works for real AI workloads. Titan #2 has the manufacturing precision to produce these chips at scale. Titan #3 has eliminated the data-transfer bottleneck that has “plagued AI chips from the start.” The convergence of these three, Gilder argues, represents “the beginning of the end for today’s big data centers” and an opportunity for “generational wealth.”

Beyond the headline thesis, Gilder bundles three “special report” bonus picks: a company benefiting from the Trump administration’s 15% revenue-sharing deal on AI chip exports to China, a “sleeping giant” ASIC designer that rivals the industry leader on performance-per-watt, and a “picks and shovels” semiconductor equipment company that every advanced chip must pass through.

The Mechanism

The three Titans map to three public companies, though the fit gets loose in places.

Titan #1: the wafer-scale pioneer (AMD or Tesla)

This is the weakest match in the lineup. Gilder provides almost no specific clues: “Titan #1’s wafer-scale design creates single-chip supercomputers that process data at unprecedented rates.” StockGumshoe’s Thinkolator narrows it to either AMD or Tesla. AMD builds “chiplet” architectures that pack multiple processing dies onto a single substrate, approaching wafer-scale integration without being a true single-wafer design. Tesla’s Dojo training system initially used a full-wafer approach with 25 processors on a single wafer before the company abandoned the design in favor of its AI6 chip, which uses conventional packaging.

If Gilder wrote the ad before last summer, he was almost certainly pitching Tesla. If he updated it, AMD is the more plausible target. AMD has invested in Cerebras, participates in its funding rounds, and the two are seen as “frenemies” working to break NVIDIA’s near-monopoly on advanced AI training chips. Regardless of which company Gilder actually means, AMD is the investable version of this thesis: it trades at roughly 60X trailing earnings, with analysts projecting roughly 80% earnings growth in both 2026 and 2027 as its GPUs and CPUs ride the AI server boom. On 2028 estimates, the forward multiple drops to about 29X. For context, NVIDIA trades at 32X trailing and 13X forward on 2028 estimates.

The risk is that AMD’s AI GPU market share remains small relative to NVIDIA’s estimated 80%+ dominance, and the company’s growth narrative depends heavily on Intel continuing to stumble in the data center CPU market.

Titan #2: the manufacturer (Taiwan Semiconductor, TSM)

This one is unambiguous. TSM manufactures essentially every advanced AI chip on the planet: NVIDIA’s H100/H200/B200 GPUs, AMD’s MI300X accelerators, and Cerebras’ Wafer-Scale Engines. At roughly $2 trillion in market capitalization and 20X forward earnings, TSM is arguably the cheapest mega-cap technology hardware company in the world. The stock has historically traded at a discount because of the ever-present geopolitical risk (China’s claims over Taiwan), but that discount has narrowed as investors have concluded that TSM is, as StockGumshoe puts it, “the one genuinely irreplaceable technology company in the world.”

TSM’s financials are formidable. Revenue grew 40% year-over-year in the most recent quarter. Advanced process nodes (3nm and 5nm) now account for over half of wafer revenue. The company is spending aggressively on global expansion (new fabs in Arizona, Japan, and Germany) to address both customer demand and geopolitical diversification. The primary risk is that semiconductor manufacturing has always been cyclical, and the current boom will not last forever. When overcapacity eventually hits, TSM’s margins and valuation will compress just as they have in every prior cycle.

Titan #3: the bottleneck eliminator (Cerebras Systems, CBRS)

This is the pure-play wafer-scale bet and the one that has captured the most attention since its May 2026 IPO. Cerebras builds what it calls Wafer-Scale Engines, literally entire silicon wafers used as single processors. The company’s CS-3 system, powered by the WSE-3 chip with 4 trillion transistors and 900,000 AI-optimized cores, is designed to train and run inference on the largest AI models without the inter-chip communication overhead that constrains GPU clusters.

The growth story is compelling on paper. Revenue was roughly $290 million in 2024 and is expected to nearly triple to $860 million in 2026. Analysts project another tripling in both 2027 and 2028, with the company potentially reaching profitability next year. At roughly $200 per share, Cerebras trades at about 80X current sales and 35X estimated 2028 earnings.

But there are significant caveats. The first is customer concentration: 85-90% of revenue comes from just two UAE-based entities, G42 and MBZ University. OpenAI and Amazon Web Services have recently become customers, which could diversify the base, but they are early in their deployments. The second is that wafer-scale computing, while elegant in theory, is unproven at the scale that would threaten NVIDIA. Cerebras has shipped systems to a handful of supercomputing customers and cloud providers, but its installed base is tiny compared to the hundreds of thousands of NVIDIA GPUs deployed across every major cloud and enterprise.

Bonus #1: NVIDIA (NVDA), the China ploy

Gilder’s first bonus pick describes an American technology company that struck a deal allowing it to resume selling advanced processors to China in exchange for a 15% cut to the U.S. Treasury. This is NVIDIA, and the deal is real: shipments of H200 chips to China began in July 2026, albeit in limited quantities. The China revenue is not transformative for a company doing over $100 billion annually, but the geopolitical signal matters. NVIDIA has successfully positioned itself as a “national asset” whose interests align with U.S. strategic objectives. At 32X trailing earnings and 20X forward, NVIDIA’s valuation has compressed dramatically as its earnings have soared, making it look almost reasonable by semiconductor standards.

Bonus #2: Broadcom (AVGO), the ASIC giant

The clues point squarely at Broadcom: partnerships with OpenAI, Microsoft, Meta, Google, and Amazon, custom AI accelerators that “rival the industry leader’s flagship chips,” and a valuation that Gilder’s ad claims is 45X forward earnings. That last number is wrong. Broadcom currently trades at about 25X forward earnings, and the “industry leader” (NVIDIA) is at 20X, not the 75X Gilder cites. Trailing P/E is 64X for Broadcom. The ASIC thesis is genuine: companies like Alphabet (TPUs), Amazon (Trainium), and OpenAI are designing their own AI chips and hiring Broadcom to build them because they want to reduce dependence on NVIDIA. We’ve covered Broadcom’s central role in the AI boom in more depth. Revenue grew 40% year-over-year in the most recent quarter, driven almost entirely by AI-related semiconductor sales.

Bonus #3: Nova (NVMI), the inspection gatekeeper

Nova is an Israeli semiconductor metrology company. It makes the precision inspection equipment that checks every advanced chip for defects during manufacturing. Gilder’s ad describes it as a “$2 billion company,” but that data is from 2024; the market cap is now roughly $14 billion. Revenue grew 40% year-over-year in the most recent quarter. At 40X forward earnings, it is not cheap, but if the thesis that every 2nm-and-below chip requires Nova’s inspection tools holds, the growth runway is long.

The Real Picks

Prices are approximate as of mid-August 2026. The tease prices and percent changes are from the StockGumshoe article dated July 21, 2026, and may not reflect intervening price movements.

Does the Math Check Out?

The wafer-scale efficiency claim. Cerebras’ own numbers claim that a single WSE-3 chip can do the work of dozens or hundreds of GPUs on certain workloads, primarily because it eliminates inter-chip communication overhead. This is directionally true for specific use cases (especially inference and training of very large models that do not fit on a single GPU), but it is not true across the board. NVIDIA’s H200 and B200 GPUs use NVLink and NVSwitch to achieve similar scaling with more flexibility. The 90% energy savings claim is aspirational, not demonstrated at commercial scale.

The “data center obsolescence” claim. This is where Gilder’s argument overreaches. Even if wafer-scale chips become the dominant architecture for AI workloads, the total demand for computing is growing faster than efficiency gains. This is Jevons paradox in action: as computing gets cheaper, we use more of it. The hyperscale data centers being built today will not become obsolete; they will simply be retrofitted with whatever chip architecture wins. Microsoft, Amazon, and Google are not going to abandon $50 billion data center investments because a more efficient chip exists. They will buy the more efficient chips and put them in the same buildings.

The valuation data is stale. The ad claims Broadcom trades at 45X forward earnings (it is 25X) and that NVIDIA is at 75X (it is 20X). It describes Nova as a “$2 billion company” (it is $14 billion). These errors are not trivial; they change the investment case significantly. At the actual valuations, several of these stocks are far more reasonably priced than the ad suggests.

The customer concentration risk for Cerebras is extreme. When 85-90% of revenue comes from two customers in one country, a single contract loss or geopolitical shift could devastate the business. OpenAI and AWS are promising new customers, but their commitments are early-stage and unproven at scale.

What They Got Right

Wafer-scale computing is a genuine technological innovation. Cerebras has shipped working systems to real customers, including supercomputing centers and cloud providers. The WSE-3 chip is not vaporware; it exists, it runs, and it delivers impressive performance on specific AI workloads. This is a real company with real technology, not a PowerPoint pitch.

Taiwan Semiconductor’s irreplaceable position is accurately described. Every advanced AI chip on the market flows through TSM’s fabs. Samsung and Intel are years behind in process technology and volume. Until a credible alternative emerges, TSM’s pricing power and growth trajectory remain intact.

The “diversification away from NVIDIA” trend is accelerating. Alphabet (TPUs), Amazon (Trainium), Microsoft (Maia), and OpenAI are all designing custom AI chips, and Broadcom is the primary beneficiary. The ASIC market is real and growing; Broadcom’s AI revenue grew 40% year-over-year in the most recent quarter.

The semiconductor equipment thesis has structural tailwinds. As chips shrink to 2nm and below, inspection and metrology become more critical, not less. Companies like Nova that provide essential manufacturing tools benefit from both increasing chip complexity and expanding fab capacity.

What They Got Wrong

The valuation numbers in the ad are wrong by wide margins. Claiming Broadcom is at 45X forward (actual: 25X) and NVIDIA at 75X (actual: 20X) is not a rounding error. It paints a misleading picture of relative value. Investors reading the ad would conclude Broadcom is significantly cheaper than NVIDIA; the opposite is true.

“Titan #1” is too vague to be actionable. If Gilder cannot clearly name the wafer-scale pioneer, it is either Tesla (which abandoned its wafer-scale project) or AMD (which does not make true wafer-scale chips). Neither is a clean fit for the thesis, which makes the entire “Titan” framing feel forced.

Wafer-scale chips will not make data centers obsolete. Computing demand grows faster than efficiency. Every prior efficiency breakthrough, from mainframes to PCs to cloud computing, increased total computing consumption, not decreased it. The data centers being built today are multi-decade assets that will host whatever chip architecture wins.

The Cerebras customer concentration problem is glossed over. An investment in CBRS today is essentially a bet that two UAE entities will continue to write large checks and that OpenAI/AWS will eventually replace them as primary customers. That is a lot of dependency risk for a company trading at 80X sales.

The ad relies on a fear-based narrative that does not hold up to scrutiny. “AI data centers are write-offs waiting to happen” is an attention-grabbing line, but the actual mechanism is gradual displacement, not catastrophic obsolescence. Infrastructure investors are not idiots; they are modeling efficiency improvements into their capex decisions.

The Verdict

Gilder has assembled an interesting portfolio of semiconductor companies under a dramatic narrative, but the narrative oversells the technology and the portfolio is available at better valuations than the ad suggests. TSM at 20X forward earnings is the most defensible pick; it is genuinely irreplaceable, reasonably valued, and benefits from every AI chip architecture, not just wafer-scale. Cerebras is the most speculative: a real company with real technology, but extreme customer concentration and an unproven market make it a high-risk bet at 80X sales.

If you want exposure to the wafer-scale thesis, TSM is the safest way to get it; they manufacture the chips regardless of who designs them. Broadcom offers a different angle: the ASIC trend that reduces dependence on NVIDIA, the same dynamic we traced through Broadcom’s optical-chip roadmap. And if you believe AI computing demand will keep growing faster than efficiency gains, NVIDIA remains the most direct play, now trading at just 20X forward earnings after its dramatic multiple compression.

The promo’s urgency is manufactured (wafer-scale computing is a multi-year technology transition, not an imminent disruption), but the underlying semiconductor themes are real. Just don’t pay attention to the valuation numbers in the ad.

This is not financial advice. NewsletterVetter has no position in any stock mentioned. Past performance, including the July 21 tease-to-close changes shown above, does not guarantee future results. Technology investments carry significant risk, and the semiconductor industry is historically cyclical.