The world for technology changed last week with the announcement of President Trump’s “Liberation Day” of massive, ill-considered chainsaw-like tariffs. They could end US dominance in AI in one fell swoop.

But before Trump’s “Liberation Day” became “Obliteration Day” , the tech world spent last week staring into fiberoptic transceivers, a REALLY important research note on that below. This week we keep an eye on TikTok, with the same geopolitical rationale and the same unanswered technical questions—about how data flows, who controls it, and what “control” means when information is more fluid than hardware. At the same time, tariffs are quietly distorting the economics of infrastructure, with effects most visible in optical components, power systems, and the companies building AI networks. The through-line is interconnection—between countries, between clusters, and between the vague rhetoric of security and the hard cost of doing business.

As always, I’m focused on three things:
1) Technology-driven change;
2) the latest in innovation and startup trends, and;
3) stock fraud.

Companies Discussed

In This Note:

• Companies Discussed

• US AI’s Death By Tariff

• The Golden Ratio That Rules AI

• Tweet O’ The Week

• Epistrophy In The News

• 📆 of Epistrophy Events

• Availability This Week

US AI’s Death By Tariff

The U.S. has long led in artificial intelligence thanks to its dominant semiconductor firms and massive cloud infrastructure. That may have ended last week. President Trump’s tariffs – including a baseline 10% duty on all imports (rising to 54% on Chinese goods) – could end US dominance in AI.

Here’s how.

These sweeping trade barriers will squeeze U.S. tech giants on two fronts:

(1) cutting off lucrative markets and critical suppliers, and;

(2) inflating the cost and slowing the pace of new data center buildouts that power AI. Below we present primary data showing U.S. chipmakers’ China exposure and detail major data center projects now at risk. The tariffs self-sabotage America’s AI lead.

U.S. Chipmakers’ China Exposure: Revenue at Risk

Many top U.S. semiconductor and mega-cap tech firms derive a significant share of sales from China – sales now imperiled by tariffs and geopolitical tensions. Recent SEC 10-K filings reveal just how dependent these companies are on the Chinese market and how concerned they are about trade restrictions:

NVIDIA (NVDA:NASDAQ): The GPU leader generated $17.1 billion (13%) of its FY2025 revenue from China. Nvidia explicitly warns that geopolitical tensions or export controls involving China “could have a material adverse impact.” Less sales weakens Nvidia’s position — and weakens US AI.

Intel (INTC:NASDAQ): Intel calls China “one of our largest markets.” In 2024, 29% of Intel’s revenue came from China – roughly $15.5 billion. Intel warns tariffs “can increase our manufacturing costs… [and] limit our ability to sell.” Even before these massive tariffs, Intel said such tensions have already affected customer orders last year. Saving Intel is critical to US AI. With 76% of its revenue overseas, Intel is highly vulnerable.

Qualcomm (QCOM:NASDAQ): San Diego-based Qualcomm is especially China-dependent, with 66% of FY2024 revenue from China. It warns: “A significant portion of our business is concentrated in China…” Qualcomm also notes that tariffs on chips or devices using its chips would harm sales.

Broadcom (AVGO:NASDAQ): Broadcom derives 20–35% of its revenue from China. The company has warned of “increased tariffs that harm our ability to participate in Chinese markets or compete.” With tariffs now rising to 54% on Chinese imports, Broadcom faces higher costs and potential retaliation and an end of it’s amazing growth story.

Amazon (AMZN:NASDAQ): While not heavily exposed to Chinese revenue, Amazon depends on Chinese manufacturing. Tariffs could add billions in landed inventory costs. As much as $7 billion in gross margin could go up in smoke, crippling AI spending.

CoreWeave (CRWV:NASDAQ): The recently-public AI cloud firm is heavily reliant on Nvidia. And it’s investing over $1  billion into new U.S. data centers. With Microsoft (MSFT:NASDAQ) as its largest customer (62% of revenue), any supply chain hiccup would affect not just CoreWeave but the broader AI ecosystem.

Taken together, these data show billions in U.S. tech revenue now at risk – revenue that funds R&D and keeps American firms on the cutting edge. Losing the Chinese market could mean less scale and capital for U.S. companies to invest in next-generation AI. As Qualcomm notes, Chinese OEMs are already developing their own chips out of fear of U.S. restrictions.

Major U.S. Data Center Projects – and Rising Construction Costs

Tariffs strike at the physical backbone of AI leadership: data centers. These projects now face delays and budget overruns from rising material costs.

Meta (META:NASDAQ)’s Temple, Texas Campus: A $800M, 900,000-sq-ft hyperscale project. Tariffs include a 25% duty on imported steel, which will increase construction costs per square foot.

Google (GOOG:NASDAQ)’s Kansas City & Iowa Facilities: A $1B campus in Missouri and $576M project in Iowa. Projects may suffer from rising prices for imported cooling and electrical equipment.

AWS’s Virginia Expansion: A $35B buildout across new campuses through 2040. Tariffs could hit power delivery infrastructure, networking gear, batteries, and GPUs, while China’s rare earth export controls exacerbate the risk.

How Tariffs Hamper AI Infrastructure

Tariffs on imported construction materials, particularly steel, are driving up the cost of building new data centers—facilities that form the physical backbone of AI infrastructure. Under the April 2025 trade policy, imported steel faces a 25% tariff layered atop the baseline 10% duty applied to all imports. Given that steel makes up an estimated 15–20% of total data center construction costs, this effectively raises the cost per square foot of new builds by as much as 8–12%, depending on the scope and material mix. For an $800 million hyperscale campus like Meta’s facility in Temple, Texas, that could translate to $60–100 million in added construction expense, potentially forcing developers to delay, downsize, or reallocate budgets away from high-density AI compute. Structural steel, racks, HVAC support frames, and shielding all require imported components—none of which can be easily or cheaply substituted with domestic supply in the near term.

Meanwhile, the digital infrastructure that powers AI workloads—servers, networking gear, and electrical systems—is also being hit. Imported power electronics and networking components, many of which contain copper or high-efficiency transformers, are now subject to tariffs ranging from 10% to 54%, depending on country of origin. These costs flow through to mission-critical equipment like switchgear, transformers, and backup generators, which are already under supply strain due to long lead times and surging demand. Major utility transformers now take over two years to procure. With copper also subject to tariff-driven price volatility, the result is not only higher costs but also a tighter bottleneck in grid connectivity for new data center projects. This is especially problematic in power-constrained markets like Northern Virginia, where hundreds of megawatts in AI infrastructure are currently stalled awaiting upgrades. Tariffs, in effect, are compounding preexisting grid delays—adding financial and logistical friction just as demand for AI compute accelerates.

Rare Earth and Raw Material Risks

The AI supply chain begins not with chips but with minerals—and China’s grip on key raw materials poses a structural risk to U.S. technology leadership. China controls more than 70% of global rare earth mining and nearly 90% of refining capacity, including production of neodymium, praseodymium, gallium, and dysprosium—critical for chip manufacturing, lasers, sensors, and high-performance magnets. In April 2025, Beijing expanded export restrictions on gallium and rare earth alloys, raising the stakes for companies dependent on these inputs. Coherent (COHR:NYSE), which manufactures precision optical components and lasers for AI data centers, defense, and medical systems, relies on rare-earth doped glass and phosphors sourced from China. Fabrinet (FN:NYSE), a contract manufacturer for optical transceivers and photonics systems, uses Chinese-sourced yttrium and other materials in high-speed components essential for AI networking. Even Rockley Photonics, a lesser-known silicon photonics company developing biosensing chips and optical interconnects, is vulnerable to gallium shortages affecting wafer substrates and photonic modulators.

Lithium is another chokepoint. China refines more than 60% of the world’s lithium and dominates production of cathode materials and electrolyte precursors for lithium-ion batteries. These batteries power not just electric vehicles but also backup power systems in hyperscale data centers. Vertiv (VRT:NYSE), which supplies power and thermal infrastructure to Google, Meta, and Microsoft, integrates lithium-based energy storage into its UPS systems and would face higher costs or delays if lithium or cobalt-based inputs are disrupted. Bloom Energy (BE:NYSE), which provides solid oxide fuel cells increasingly used as auxiliary power in data centers and microgrids, also sources materials from China’s tightly controlled supply base. Even Generac (GNRC:NYSE), best known for residential generators, is now marketing lithium-based storage to commercial and industrial customers—tying its growth to fragile energy material supply lines. As trade restrictions bite, these companies could see increased costs, longer lead times, and tighter constraints on scaling their AI-adjacent infrastructure offerings. The result isn’t just pressure on headline tech firms but on the deeper industrial base that makes AI infrastructure possible.

Undermining U.S. AI Leadership

If these tariffs persist, the U.S. could lose its lead in AI by undermining the global scale, agile supply chains, and infrastructure capacity that gave it an edge. As the filings and project data above show, Trump’s April 2025 tariffs immediately disrupt chip earnings and AI infrastructure deployment. With a stroke of the pen, President Trump has crippled US AI so early in this AI race.

The Golden Ratio That Rules AI

Euclid’s Golden Ratio—1.618:1—defined proportion for the ancients. But in artificial intelligence, there’s a new AI Golden Ratio that matters more: how many interconnects are needed for every AI cluster. It was 2:1 two years ago. It’s 5:1 today. By 2028, it’ll be 10:1. And unlike Euclid’s, this ratio isn’t eternal—it’s accelerating. For investors who understand what that means, it could well be the closest thing to found gold.

An AI cluster is a tightly coupled array of GPUs or custom accelerators—Nvidia (NVDA: NASDAQ) B100s, Google (GOOGL: NASDAQ) TPUs, Amazon (AMZN: NASDAQ) Trainium—designed to operate as a single unit. These aren’t standalone chips. They function as a system through high-speed, low-latency interconnects that link processors across nodes and racks. Without interconnects, there is no cluster—only loose compute. Copper cabling served well in the early days of AI, when models were smaller and latency budgets were forgiving. But as architectures scaled up, copper hit hard limits: signal degradation, energy cost, and reach. Optical interconnects—based on silicon photonics and fiber—don’t just extend those limits. They replace them.

At the OFC 2025 conference in San Francisco, I asked engineers, suppliers, and network architects a simple question: What ratio of interconnects to AI clusters are you planning for? They didn’t all agree on a number. Some said 8:1, others said 10:1 or more. But nobody said it was going down. The consensus wasn’t about the precise figure—it was about the slope. Interconnect demand is growing faster than cluster size. And that slope is steepening.

Today’s working ratio is five interconnects per AI cluster. That’s up from two in 2022. The trend shows up clearly in deployments like OpenAI’s GPT-4 training cluster, which used roughly 25,000 GPUs connected by more than 75,000 optical links. Or Meta’s AI Research SuperCluster, which in its 2022 build integrated 6,080 NVIDIA A100s and over 15,000 InfiniBand links. Tesla’s Dojo system, though proprietary, is another outlier example: each “tile” of 25 chips is highly interconnected, with over 9 Tb/s of internal bandwidth per tile. The design assumes more communication, not less, as scale increases.

There are two ways to make a cluster bigger: “scale up” (stack more GPUs in a single system) or “scale out” (add more systems and connect them across an optical fabric). In practice, both are happening. But scale-out architectures will come to dominate for one simple reason: AI training jobs no longer fit on a single machine. As clusters grow, interconnects grow faster. Bigger clusters require more internal links. Distributed clusters require external connections that multiply geometrically with each node added. A 100,000-GPU build isn’t just four times a 25,000-GPU system—it’s ten times the network.

And copper can’t make that leap. At Nvidia’s GTC in March, Jensen Huang said it plainly: “copper, copper, as far as we can.” The message: they’ve hit the ceiling. Beyond five meters, copper cabling struggles with power draw, electromagnetic interference, and thermal load. Optical interconnects solve those problems. Silicon photonics consumes 5 pJ/bit versus 15–25 pJ/bit for copper. And with co-packaged optics reaching 1.6 Tb/s per port, they scale more gracefully as bandwidth demand increases.

That demand is no longer speculative. Cluster sizes are already on a visible trajectory:

Based on my research this week, I expect the ratio itself—now 5:1—will likely double again within three years. That’s not a rounding error. It’s an architecture shift. You can no longer build compute without planning the network at the same time. And for some builders—cloud hyperscalers, LLM developers, chip designers—the network has become the primary constraint.

Why will scale-out keep winning? Three reasons.

First, power density: high-end GPUs consume 700W or more. Put enough of them in a rack and you hit thermal limits faster than you hit compute limits. Spreading workloads across multiple systems avoids hot spots and reduces cooling costs.

Second, fault tolerance: distributed clusters degrade gracefully. If a rack or node fails, training can reroute without collapsing. Scale-up systems are brittle. Fail a board, lose the run.

Third, geographic flexibility: large training jobs increasingly span data centers. Companies want to deploy across regions or availability zones for redundancy, latency shaping, or regulatory reasons. That demands high-bandwidth links between buildings, not just racks. Only optical survives that stretch.

The result is a compounding equation: as clusters scale out, the interconnect count rises faster than the GPU count. The Golden Ratio—interconnects per cluster—isn’t just a diagnostic. It’s a forecast. Every uptick favors optical, and every jump benefits a specific supply chain. At the top are optical component makers like Applied Optoelectronics (AAOI: NASDAQ), Coherent (COHR: NYSE), and Lumentum (LITE: NASDAQ), whose lasers, modulators, and photonic integrated circuits drive signal throughput. Contract manufacturers like Fabrinet (FN: NYSE) assemble and package those components into transceivers and modules. Ribbon Communications (RBBN: NASDAQ), long a specialist in optical transport, is repositioning itself as a key vendor for hyperscale interconnects.

What I saw at OFC this year wasn’t a debate—it was a realignment. Everyone now assumes the future is optical. The conversation has moved on from whether to when, and from bits to power. Lasers are being co-packaged with switches. Silicon photonics is entering its third generation. And yet, the demand curve is still outpacing the roadmap.

The Golden Ratio of AI is now 5:1. Within two years, it will be 10:1. That number, invisible to most, is the engine behind everything else. It explains why network topologies are being redesigned, why new protocols are emerging, and why old assumptions about data center architecture are breaking down. Ratios don’t usually make headlines. This one should.

Tweet O’ The Week

Epistrophy In The News

I joined Fox Business’ Charles Payne, an old friend, for the first time in a long time, as the market suffered its biggest one day drop since the Pandemic. On NewsNation, we threw out the script, parsing the TikTok bill as a network problem, not just a national security one—and had plenty of talk about the impact of terrifs. Then, ninety minutes with Julie Hyman on Yahoo Finance, where the conversation ranged from semiconductors to transpacific supply chains and, of course, enterprise AI spending.

📆 of Epistrophy Events

Availability This Week

Back in New York early, then San Francisco later this week, mostly working out of the Ferry Building, of course. As always, happy to meet with journalists, investors, or policy folks who want to talk shop—optics, chips, AI, or anything else that's quietly reshaping the stack.

Written reports are available to clients, with video summaries on YouTube, and of course our popular summaries of the summaries on Instagram, TikTok, and YouTube Shorts.

I hope these notes are helpful to you. I’d love to discuss them further. If you have a friend or even a frenemy whom you think might benefit from this note, have them reach out and I’ll put them on the list.