Earnings season returns to technology. Vertiv (VRT: NYSE), IBM (IBM: NYSE), Tesla (TSLA: NASDAQ) and Intel (INTC: NASDAQ) all report this week—four companies that together mark the pulse of AI infrastructure, from power and cooling to chips and servers to demand itself. Their results will hint at whether the AI buildout has legs beyond hype, and whether the capital behind it is still multiplying.

Meanwhile, last week’s OpenAI–Broadcom (AVGO: NASDAQ) deal and Oracle’s (ORCL: NYSE) AI World underscored the deepening tangle between hardware, software and sovereignty—who controls what and where it’s built.

IIn the midst of this, Epistrophy wrapped up another major study, this time focused on Dark Fiber (sounds scary, right?) We’re hearing more and more how data centers will be connected by the thousands of miles of optical fiber laid in the dot com era. But by tracing hundreds of pages of SEC filings, legal complaints, technical documents, construction permits and specifications going back 25 years, we concluded that 42% of America’s “dark fiber” is useless and cannot support modern AI backbones. It’s a problem. And at the same time it’s a huge opportunity for a small handful of companies, whom we identify.

Check out our report below.

Our website is tight! Check out our repository of past notes and our searchable research database at epistrophy.beehiiv.com.

As always, we’re 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

• Spire’s Spacey ETF Trip 🚀 

  • An ETF’s Gains Create a Space Bubble

• Negative Infrastructure

  • The Big Dig

  • Why Dark Fiber Can’t Power AI

  • Weaving Racks: Why Scale Out Is the Only Way

  • Negative Infrastructure: What’s Really in the Grou …

  • The New Silk Road

  • The Path To Abundance

• Tweet O’ The Week

• Epistrophy In The News

• 📆 of Epistrophy Events

• Availability This Week

Spire’s Spacey ETF Trip 🚀 

An ETF’s Gains Create a Space Bubble

In early October 2025, Spire Global (SPIR: NASDAQ) saw its share price rise from $10.87 on October 2 to $14.40 on October 9. An honest to God 🚀. But why?

This could be a bizarre case where business failures are leading to a rise in the stock price. 

Spire competitor Rocket Lab (RKLB: NASDAQ) has had good news. Its offerings of dedicated launch services and satellite deployment capabilities overlap with Spire’s mission of providing space-based data and analytics to commercial and government clients. Rocket Lab announced multi-launch contracts in late September and early October – contracts that eluded Spire. 

The deals received significant media coverage, driving a sharp increase in the shares of Rocket Lab — and the Procure Space ETF (UFO: NASDAQ), of which Rocket Lab is a top holding.

And even though Spire did not get the contracts, the ETF was compelled to buy more share of Rocket Lab and Spire – as well as the other space and satellite equities in the ETF. That led to an outsize requirement to buy more Spire. 

Empirical research on ETF effects has garnered recent attention from academics, including “ETF effects: The role of primary versus secondary market activities,” “An empirical analysis of exchange-traded funds in the US”, and “Value of ETF Liquidity (The Review of Financial Studies)”. They’ve demonstrated that ETF inflows amplify price movement for basket components—regardless of whether those companies generate news. ETF trading algorithms and authorized participants typically rebalance by buying shares across all holdings during periods of heavy sector inflows, mechanically driving prices higher for names such as Spire Global, even when company fundamentals do not support the move.

Assets under management in UFO approached $149 million in October 2025, and trading volumes spiked after Rocket Lab's news, consistent with patterns identified in academic research. In the context of concentrated funds and limited liquidity in space equities, these ETF-induced swings can be material and abrupt.​  Spire’s daily traded volume spiked above its average, reaching as high as 774,918 shares on October 8 compared to an average volume of around 557,000 shares, translating to daily cash turnover in the millions of dollars. Such volume surges illustrate how ETF-driven activity can make trading in Spire abrupt and substantial.

Academic studies provide extensive models for price response, indicating that a single large move by a top ETF holding can trigger comparable, short-lived rallies across other constituents—regardless of underlying operational or financial health.

Short-lived price gains tend to be just that. If Spire Global reports another lousy quarter in November,, ETF-boosted gains may be vaporized just like space dust that drove the rally in the first place.

Negative Infrastructure

Our extensive study of Dark Fiber, its uselessness for the coming AI Buildout and 24 companies that could benefit.

In Richland Parish, Louisiana, a $10 billion machine is rising from the dirt. Meta Platforms (META: NASDAQ) is building a four-million-square-foot campus , roughly the size of Disneyland or larger than the entire footprint of midtown Manhattan south of 42nd Street.  It’s Meta’s largest data center yet: a $10 billion, four-million-sq-ft campus designed to deliver over two gigawatts of compute and employ 5,000 skilled trades in its construction. 

Hundreds of thousands of GPUs will be linked together by substations, transformers, chilled water and above all, optical fiber. The fiber will link this massive data center to other data centers hundreds of miles away to tackle unimaginable levels of simultaneous compute.

That should be good news: America already has billions of dollars of fiber in the ground. Most of it is dark, unused since the telecom bubble. On paper it looks like surplus. 

But the dirty secret is that billions of dollars of that dark fiber is stranded. It was built for another era, another traffic model — long-haul client/server flows — not the synchronous east–west patterns of AI. 

We’ve undertaken an extensive study of just what’s in the ground, why so much of it is bad and who the beneficiaries of the next buildout might be. Sifting through the forensic record of the last telecom bubble—SEC complaints, bankruptcy schedules, and sworn testimony from WorldCom, Global Crossing, and Qwest—cross-referenced against Corning’s fiber specifications and Level 3’s engineering disclosures. Layered with modern IEEE/ITU optical standards and IX facility reports from Equinix ( EQIX: NASDAQ ) and Digital Realty ( DLR: NYSE ), this archive makes clear which strands of buried glass can still carry 400–800 G and what’s dead in the ground – dark fiber we’re calling “negative infrastructure.”

The Big Dig

The late 1990s were a mania for glass. Global Crossing laid undersea cables at breakneck speed, linking continents. Level 3 Communications dug conduits along railroads, later absorbed into Lumen Technologies (LUMN: NYSE). Qwest (also folded into Lumen) raced to build its own backbone. Enron launched a broadband unit that promised to trade bandwidth like oil.

Executives acted as if every mile of fiber dug would be worth more tomorrow. Global Crossing threw parties aboard yachts in Bermuda to celebrate new cables. Enron Broadband staged an infamous analyst day in 2000 where executives faked a video stream to demonstrate capacity that didn’t exist. Qwest engineered “indefeasible rights of use” swaps, trading dark fiber with other carriers to book fictitious revenue. WorldCom took it further, committing $11 billion in accounting fraud tied in part to capacity swaps.

The money was staggering. Between 1998 and 2001, more than $100 billion went into U.S. long-haul fiber. Then bandwidth prices collapsed by 90%. Carriers wrote down billions. Investors were wiped out.

The fiber was still in the ground, but the market treated it as a liability. Every mile dug at peak valuations weighed down balance sheets when demand failed to appear. Dozens of companies collapsed, not only because the demand wasn’t there, but because their executives had monetized growth itself. Stock options and bonuses rewarded route-miles and new ducts, not free cash flow. The result was ruin — and fraud.

Why Dark Fiber Can’t Power AI

The strands buried in that boom still exist. Maps show conduits. Manholes still hold coils. But they cannot meet the demands of AI.

Wrong geography. The glut was long-haul. It connected St. Louis to Denver, Dallas to Los Angeles. AI requires dense metro rings, low-latency loops that connect multiple campuses in Ashburn, Dallas, and Phoenix. Latency budgets are measured in microseconds. A strand across Kansas does nothing for two data halls in Virginia that must act as one.

Wrong optical budget. Much of the 1990s glass was specified for 2.5 or 10 Gbps. Modern coherent optics from Ciena (CIEN: NYSE), Coherent (COHR: NYSE), Lumentum (LITE: NASDAQ), and Cisco Systems (CSCO: NASDAQ) now push 400 to 800 Gbps. Old strands suffer attenuation and chromatic dispersion that make such speeds uneconomical without regeneration — which costs more than trenching new.

Wrong architecture. Legacy networks assumed static circuits: one port, one line, one bill. AI requires elastic fabrics that can reallocate terabits instantly. The old systems cannot adapt.

Wrong economics. Even if strands could be lit, rights-of-way still incur taxes and permits. Many conduits have been flooded, cut, or abandoned. Maintenance costs climb. In many cases, it is cheaper to build fresh than to revive the past.

The surplus is not capacity. It is a ghost of capacity.

Dark, Dead And Buried

Based on our review of SEC filings, bankruptcy records, vendor disclosures, and technical standards, we found that 42% of the dark fiber still buried under American soil is effectively unusable for modern AI backbones. The number comes not from guesswork but from the history of how different generations of fiber were specified, deployed, and later abandoned.

The first large wave of fiber buildouts in the late 1980s and early 1990s used G.652 A/B/C, standard single-mode fiber. It worked well for the 1310 nm transmission window, but it carried a serious defect: a “water peak” around 1383 nm. 

Dispersion-Shifted Fiber (DSF, G.653) looked like an innovation in the 1990s: by shifting the fiber’s zero-dispersion point to 1550 nm, it matched the sweet spot of optical amplifiers. But when dense WDM arrived, the lack of dispersion caused channels to mix with each other, creating four-wave mixing and crosstalk. What was once ideal for single-channel links became almost useless for modern coherent optics — stranding nearly a fifth of the dark fiber base.

Older G.652 fibers absorbed light near 1383 nm — the “water peak” — blocking a crucial part of the spectrum and reducing how many WDM channels could be used. That absorption spike restricted the usable spectrum, limiting the number of dense Wavelength-Division Multiplexing (WDM) channels that could be lit in later decades (WDM allows dozens of light signals, each at a different wavelength, to travel simultaneously down the same strand of fiber.) These early G.652 strands still exist in metro and access networks, and while they can be coaxed into carrying 400 G coherent signals with margin penalties, they are generally brittle at 800 G. They account for 23% of the remaining dark inventory.

By the mid-1990s, long-haul builders like Global Crossing and WorldCom adopted G.653 dispersion-shifted fiber. It was designed to eliminate dispersion at 1550 nm, which seemed a perfect match for the optics of the time. The flaw was that near-zero dispersion created catastrophic nonlinearities once dense WDM systems arrived. Four-wave mixing and cross-talk made this fiber almost useless for modern coherent optics. Industry consensus is clear: legacy DSF is stranded. This category represents 19% of what is still in the ground, and it is the clearest example of negative infrastructure.

The late-bubble builders shifted course. Level 3’s disclosures show that between 2000 and 2001 it purchased more than two million cabled-kilometers of Corning’s LEAF fiber, a G.655 non-zero dispersion-shifted fiber (NZDSF). This fiber was engineered to avoid the nonlinearities of DSF while keeping dispersion manageable for long-haul DWDM. NZDSF introduced a small, controlled amount of dispersion at 1550 nm, just enough to prevent the destructive mixing that doomed DSF. Because DSF had almost zero dispersion at 1550 nm, overlapping WDM signals interfered with each other, producing four-wave mixing and crosstalk that crippled capacity under modern coherent optics. Unlike its predecessors, it remains serviceable today, supporting 400 and 800 G coherent links with appropriate amplifier spacing and PMD margins. It represents 18% of the base.

Metro construction after 2003 almost universally adopted G.652.D, the “low-water-peak” standard. Corning’s SMF-28e and AllWave families opened the full 1260–1625 nm window, enabling far denser channel plans and compatibility with coherent optics. These strands remain the healthiest of the legacy base, and make up 37% of the total.

Multimode fiber (OM1–OM4) makes up about 3% of the dark fiber base, but because it is only suited for short-reach connections inside buildings, it is irrelevant for hyperscale interconnect.

Weaving Racks: Why Scale Out Is the Only Way

The compute itself drives the network. AI training requires synchronous stochastic gradient descent. Every GPU must exchange updates with every other GPU in the cluster, each step of training. That creates all-to-all communication patterns at terabit speeds.

Power and thermal limits cap the density of a single hall. Once racks saturate those limits, the only solution is to extend the cluster across buildings, then across campuses, then across regions. Each extension multiplies the bandwidth demand.

The workloads make the case:

• Large Language Models. Training a trillion-parameter model requires tens of thousands of GPUs. Each forward and backward pass demands gradient all-reduce operations across the entire cluster, that must expand across geographies,
  

• Recommendation Engines. Meta’s embedding tables span terabytes. Training requires sharding them across multiple campuses.
  

• Multi-Modal Models. Combining video, image, and text datasets requires petabytes of input streaming across clusters.
  

• Cybersecurity & AI-powered Hacks. Safety runs replicate models across sites for red-teaming, multiplying the bandwidth again.

This is why hyperscalers treat metros as single logical machines. A campus in Oregon may act as a rack linked to one in California. A hall in Ashburn must synchronize with a hall in Columbus. The determinant is not geography but interconnect: whether the fiber can deliver deterministic latency and terabits per second across regions.

Negative Infrastructure: What’s Really in the Ground

Epistrophy Capital Research did an extensive review of twenty-seven years of SEC filings, bankruptcy records, vendor disclosures and technical standards reviewing the optical fiber laid into the ground across the continental United States. Based on those documents, we’ve determined that 42% of America’s dark fiber can’t be used for AI data centers. We’re calling that “negative infrastructure.”

Here’s the breakdown: 

The first wave of deployment in the late 1980s and early 1990s used Standard Single-Mode Fiber (SSMF, G.652 A/B/C). It carried a “water peak” around 1383 nm that restricted usable spectrum. These strands survive in metro and access networks and can sometimes handle 400 G coherent with penalties, but they are brittle at 800 G. They account for 23% of the remaining dark inventory.

By the mid-1990s, long-haul builders such as Global Crossing and WorldCom turned to Dispersion-Shifted Fiber (DSF, G.653), designed to reduce dispersion at 1550 nm. In practice, it proved catastrophic under DWDM and coherent optics because of nonlinear four-wave mixing. This fiber represents 19% of the total and is essentially stranded.

Late in the bubble, Level 3’s filings show it bought more than two million cabled-kilometers of Corning’s LEAF®, a Non-Zero Dispersion-Shifted Fiber (NZDSF, G.655) engineered to sidestep DSF’s flaws. Unlike earlier builds, it remains serviceable for 400/800 G today, representing 18% of the base.

Metro construction after 2003 largely standardized on Low Water-Peak Single-Mode Fiber (SSMF, G.652.D), which opened the full 1260–1625 nm window. These strands, now 37% of the stock, are coherent-friendly and form the healthiest share of legacy fiber.

The arithmetic is blunt: the 19% of G.653 DSF and part of the 23% of older G.652 A/B/C SSMF add up to 42% of America’s dark fiber plant that cannot economically be lit at 400 or 800 G. In practice much of this glass is negative infrastructure: sunk capital that costs more to revive than to replace.

The New Silk Road

The demand for new interconnect could lead to a new set of winners across the stack. Some of these companies will flub this once-in-a-lifetime opportunity – but some won’t. 

Companies to watch:

Glass. Corning (GLW: NYSE) remains dominant. Prysmian (PRY: BIT) and Sumitomo Electric (5802: TSE) ship reels worldwide. CommScope (COMM: NASDAQ) supplies cable systems.

Optics. Fiber matters only if lit. Ciena (CIEN: NYSE) leads in transport. Applied Optoelectronics (AAOI: NASDAQ) promises next0gen transceivers to hyperscalers. Coherent (COHR: NYSE) makes lasers. Lumentum (LITE: NASDAQ) supplies modules. Fabrinet (FN: NYSE) assembles much of this gear.

Switching & Routing. Arista Networks (ANET: NYSE) dominates data center switching. Juniper Networks (JNPR: NYSE) focuses on high-performance routing. Nokia (NOK: NYSE) remains a force in optical transport. Cisco Systems (CSCO: NASDAQ) straddles both, combining backbone routing with coherent optics through its Acacia unit.

Carriers. Lumen Technologies (LUMN: NYSE), Crown Castle (CCI: NYSE), American Tower (AMT: NYSE), AT&T (T: NYSE), and Verizon (VZ: NYSE) expand metro and regional loops.

Hyperscalers. Amazon (AMZN: NASDAQ), Alphabet (GOOGL: NASDAQ), Microsoft (MSFT: NASDAQ), and Meta Platforms (META: NASDAQ) finance private backbones, unwilling to rely solely on carriers.

Builders. Construction firms capture billions. Comfort Systems USA (FIX: NYSE) installs cooling and electrical. Quanta Services (PWR: NYSE) builds fiber and power lines. Dycom Industries (DY: NYSE) and MasTec (MTZ: NYSE) trench and bore ducts. MYR Group (MYRG: NASDAQ) handles high-voltage hookups.

The Path To Abundance

The fiber glut of 25 years ago  wasn’t just a financial fiasco, it was often incentivized fraud. Executives at Global Crossing, Qwest and WorldCom enriched themselves with stock options and bonuses for laying glass, not for generating cash. Accounting games like capacity swaps booked phantom revenue. Wall Street rewarded “route-miles” instead of free cash flow. The result was bankruptcies, fraud convictions and billions of dollars incinerated.

Dark fiber became negative infrastructure because it was built for the wrong workload. If incentives that once rewarded scale over suitability, today’s AI build risks the same fate — but for those who understand where the real bottlenecks lie, there is a bright future and fortunes to be made.

Tweet O’ The Week

Epistrophy In The News

On the corporate fraud front, I was quoted in Laura Bratton’s Yahoo! Finance story on AI’s self-investment loop—why venture-backed startups are now the biggest customers of the same companies that fund them. I called it “very troubling” because sometimes circular finance can masquerade as demand.

I also joined Schwab Network’s Next Gen with Jenny Horne to unpack the week’s dueling conferences—Oracle’s AI World in Las Vegas and Salesforce’s (CRM: NYSE) Dreamforce in San Francisco—and what they revealed about enterprise AI spending.

And, back in the Bay Area, I did a more general market hit on Fox KTVU with Gasia Michalian to discuss the Bay Area’s AI-driven boom, the success of a handful of local companies and the reaction to political comments from the Salesforce CEO.

📆 of Epistrophy Events

Availability This Week

I’m in San Francisco for all week, preparing for a wave of Q3 tech earnings and the reporting chaos that follows. If you’re in the city, coffee’s on me.

As always: 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.