StreetSignal

Is he still acting on HBO Homeland?

When did @BillAckman become so bad at investing? loves $nflx $uber and mag7. Like that worth 2%?

Does Daniel know this is very negative? Mkt closed and they hope to complete what they got?

Problem is this math was same on Friday as today within the ratio you used so not new. And now they need the demand for it?

Peter, where is Fable if you have chart as well

Perhaps as simple as buyers live higher and sellers live lower. Pretty simple stuff?

Good thing for France they banned his dumbass ideas, blatherings and general convicted thoughts. We think he should return his SPAC scam $

You sound long 😆

For all the blind bulls pointing to GPU rental prices, our pals at $NBIS now have real time rental rates on their site. The GB300 we hope is a "mistake"

📷 StreetSignal—Our DWDM History Lesson OCT 17, 2026 Who remembers what DWDM stands for—and how Patrick Nettles, then CEO of Ciena, helped introduce it to the world in the mid-1990s? Back then, Gary Smith, now Ciena’s longtime CEO, was just the head of sales, complete with that unmistakable British accent. Yesterday’s Saturday edition is Part A—the historical foundation for tomorrow’s main event. We’re going back to the Nettles and Smith era at Ciena, when DWDM (Dense Wavelength Division Multiplexing) fundamentally changed the economics of the Internet, unlocking extraordinary growth in bandwidth, traffic, and infrastructure investment. Today, in Part B, we connect the dots between the Internet buildout of the 1990s and today’s AI infrastructure supercycle. Does greater efficiency unleash even more demand, or eventually undermine the economics of the buildout? What does history tell us about the opportunities, the risks, and where we might be in this cycle? No, we aren’t Ed Zitron, George Noble, or Jim Chanos, who vividly remembers—and profited from—Enron’s infamous “capacity trading.” Nor are we Shay Boloor or Dylan Patel, the latter having made a fortune staying ahead of the curve analyzing AI. Not a doomsayer or a cheerleader. We are StreetSignal. We invest, analyze, and share the institutional memory investors need to navigate cycles. That is before everyone was a bottleneck investing guru. Yep, each cycle has them. In the 1990s it was Alberto Vilar of Amerindo Investment Advisors fame and subsequent demise ending in fraud charges. Today, you pick your genius villain. Either way, understanding what happened before is often the key to recognizing what’s happening now—and what comes next. Saturday is the history lesson. Sunday is the investment playbook. One of the most important developments of the late 1990s, and one that most investors today probably don’t remember, was Dense Wavelength Division Multiplexing, or DWDM. Until then, the conventional wisdom was that if you wanted to carry substantially more internet traffic, you needed more fiber in the ground. That meant digging trenches, laying cable, securing rights of way and spending billions. DWDM changed that equation. By transmitting multiple wavelengths, or colors, of light simultaneously through a single strand of glass, optical equipment makers could dramatically increase the amount of data carried over existing fiber. What initially might have required dozens of separate fibers could increasingly be accomplished with one. And as the technology improved, the number of wavelengths and the capacity of each kept rising. Companies like Ciena, Nortel and Lucent were at the center of this revolution. Sound familiar? We had our DeepSeek moment in AI, which briefly looked like it might be another DWDM moment. But so far, Jevons paradox—the John Sidgmore/UUNET story of today—has worked handsomely, even as each new generation of NVIDIA GPUs delivers dramatically more tokens per dollar. Think about what that meant for the telecom companies that had just spent billions building networks on the assumption that bandwidth would remain scarce. They were racing to install more physical capacity at precisely the moment technology was multiplying the capacity of the fiber already in the ground. DWDM didn’t single-handedly cause the telecom bust; there was also enormous overbuilding, too much debt and plenty of questionable demand forecasts. But it made the supply problem considerably worse. Internet traffic continued growing at extraordinary rates, yet the amount of available bandwidth grew even faster. And once the fiber was installed, the marginal cost of selling additional bandwidth was relatively low. Operators started competing aggressively on price to fill their networks. The supposed scarcity premium disappeared, bandwidth prices collapsed and billions of dollars of infrastructure investment suddenly looked considerably less attractive. The irony was that the internet ultimately became far larger than even its biggest promoters imagined. That didn’t save the investors who financed its early infrastructure. Now consider AI. We hear constantly that compute is sold out for years, that power is constrained and that demand will absorb everything the industry can possibly build. Perhaps. But DWDM is a reminder that supply isn’t simply a function of how many data centers or GPUs you install. It’s also a function of how much useful work each dollar of infrastructure can perform. Better chips, custom ASICs, model distillation, quantization, caching and more efficient inference software can dramatically increase effective compute capacity without a corresponding increase in capital spending. There are differences, of course. Power connections and physical data centers cannot be multiplied by a software upgrade, and lower inference costs may unleash enough new demand to absorb much of the efficiency improvement. But the question for investors is the same one they should have asked in 1999: What happens to the economics of supposedly sold-out infrastructure if technology allows the industry to deliver vastly more capacity than anyone anticipated? The demand can be right, the technology can change the world, and the people financing the buildout can still lose an enormous amount of money. Elon Musk alone, through SpaceX, is talking about adding 10 gigawatts of AI compute capacity in 2027. That’s potentially $600 billion in capital expenditures from ONE player, if we take him at his word. Obviously, there is a timing issue between announcing that capacity, building it, energizing it and actually turning it on. But isn’t that a bit like fiber in the late 1990s? Back then, the industry spent enormous sums building networks based on extrapolations of seemingly insatiable Internet traffic growth. And the demand was real. The Internet did change everything. The problem was that capacity ultimately arrived faster than the revenues needed to justify it, and the economics of that capacity changed dramatically along the way. Which brings us back to DWDM and our discussion of Ciena. What happens if AI has its own DWDM moment? Back then, dense wavelength division multiplexing dramatically increased the amount of traffic that could travel over existing fiber, effectively multiplying supply without requiring a proportional increase in physical infrastructure. What if a breakthrough in inference efficiency, model architecture, memory or software suddenly allows the industry to generate 5x or 10x more useful intelligence from the same installed compute? Haven’t we already seen hints of this with DeepSeek and successive generations of NVIDIA GPUs? Yes, Jevons paradox says cheaper inference should unleash more demand, and so far it has. But what if the supply curve shifts faster than the demand curve, even as demand continues exploding? That is exactly what happened in telecom. The bears today would argue open source models are just that. Now consider the economics. Musk has reportedly been able to rent compute at rates implying one- to two-year paybacks. Insane economics, and arguably only sustainable in a market where customers are panicked about securing scarce capacity. But 10 gigawatts alone, at our illustrative assumptions, could require roughly $300 billion in annual revenues somewhere at the model or application layer to justify the investment. Sure, Grok will consume some of that capacity, but how much? And what happens to those extraordinary rental economics when this much new supply arrives? Now add Microsoft, Google, Amazon, Meta and everyone else. Yes, there is some double counting, particularly where hyperscalers are renting capacity to the very model providers whose demand supposedly justifies the spending. But if the industry ultimately adds another 30 gigawatts of AI capacity, at $20–30 billion of required annual revenue per gigawatt, we are talking about $600–900 billion in incremental annual revenue needed somewhere in the ecosystem. And what happens if an AI version of DWDM arrives while all that capacity is still under construction? Suddenly, the same demand can be served with a fraction of the originally anticipated infrastructure, just as the industry prepares to bring enormous new supply online. That doesn’t mean AI “fails”. Quite the opposite. The Internet was one of the greatest technological and economic revolutions in history, yet the companies financing its infrastructure suffered some of the greatest capital destruction investors have ever witnessed. The winners were ultimately the users of cheap bandwidth, not necessarily the owners of the networks. Could AI follow the same path, where intelligence becomes ubiquitous, consumption explodes, and the economics of supplying that intelligence deteriorate? The question isn’t whether AI changes the world. We think it does. The question is whether the capital markets are confusing an extraordinary technological revolution with an equally extraordinary investment opportunity for everyone funding its infrastructure. Those are two very different things. Disclaimer: StreetSignal is an independent financial news and market commentary publication. We aggregate, curate, summarize, and comment on information from public filings, company presentations, conference calls, news sources, third-party research, and other publicly available materials. Any opinions, interpretations, or views expressed—including whether developments may be positive or negative for a company, industry, or security—reflect our editorial judgment at the time of publication and are provided solely for informational and educational purposes. StreetSignal is not a registered investment adviser or broker-dealer, and nothing contained herein constitutes investment advice, a recommendation, or a solicitation to buy or sell any security. While we believe the information presented is derived from sources considered reliable, we make no representation or warranty as to its accuracy, completeness, or timeliness. Markets, facts, and circumstances change rapidly. Unless expressly stated otherwise, StreetSignal undertakes no obligation to update, revise, or correct previously published content, opinions, estimates, or narratives in light of new information or subsequent events. What’s “most actionable” (cards with this label in StreetSignal) is inherently subjective and depends on your individual goals, risk tolerance, and circumstances. This information is for educational purposes only and should not be considered investment, legal, or financial advice. Copyright © 2026 StreetSignal. All rights reserved. Redistribution, reproduction, forwarding, or commercial use of any portion of this publication without prior written permission is prohibited.

🤡

Everyone knows the here and now case in AI. Memory is tight thru 2029, LTAs. Compute is so so tight look at rental rates. Etc For once would be nice if people would write what could go wrong “quickly”. Poke a hole in the bull case. It reminds me of a story. Robert Bugbee CEO of STNG. You would call him and he would quote LR2 rates thru the roof. Customers begging for locking in rates for long and longer durations. Supply coming online wouldn’t be fast enough. The order book was too small to offset the growing route miles. Etc etc. In a flash, it was over. IMO2020 was not a generational change. It was a moment in time in a cycle. For AI, follow the gigs being sold by NVDA, AMD, AVGO and MRVL vs the frontier models revenues. Of course add open source to it with less potential profits per token. That is where the tell will be if things suddenly slow and supply catches up with demand.

Think we have our best written historical piece yet on AI vs DWDM. Proud of the team. Comes out at 12pm EST on StreetSignal.substack.com

Vivek would like the stocks here, 300% here or 50% lower. He and most sellsiders negative software on lows remember then after bounce find the ones to pump. Reiterate my buy on….. One thing that struck me this week is Microsoft wants to do all inference w internal ASIC. They have been $nvda largest customer to date.

I own $avgo for the software biz........

Does X have coherency lessons?