Episode 70 · 2026-08-18 · 9 min

2026-08-18 — Money, Slop, and Smoke: AI's August Reckoning

Anthropic hits $6.5B in annualized revenue and adopts invisible watermarks, while Nvidia bets big on SoftBank, platforms wage war on AI slop, Microsoft quietly retreats on Excel Copilot, and satellites start watching for wildfires in real time.

Episode summary

August 18th, 2026 finds AI money, AI mess, and AI salvation arriving simultaneously. Anthropic's explosive revenue growth and its adoption of an open watermarking standard raise the question of whether enterprise dominance and regulatory compliance can reinforce each other — or whether both are more fragile than they look. Across the episode, a throughline emerges: the AI industry is now large enough to create its own problems at scale, and the solutions it reaches for — detection filters, equity moats, deprecated features, satellite eyes — reveal just how much the governance layer is still catching up to the capability layer.

Key topics

  • AI
  • Anthropic
  • Openai
  • Infrastructure

Chapters

  1. Chapter 1

    Today, August 18th, 2026: Anthropic just crossed $6.5 billion in annualized revenue and is about to start watermarking Claude's text invisibly. Nvidia drops a $1.5 billion bet on.

  2. Chapter 2

    TechCrunch reports Nvidia is putting $1.5 billion into SoftBank's data center development arm — the same arm building an OpenAI facility. That deal locks Nvidia chips in as.

  3. Chapter 3

    The New York Times has a piece on Spotify, LinkedIn, and other platforms deploying AI detection systems to filter out low-quality AI-generated content — what the piece calls.

  4. Chapter 4

    The Register reports Microsoft is deprecating the dedicated Copilot worksheet function in Excel. The side-pane interface is staying; the formula-based approach is going. Microsoft says the pane is.

  5. Chapter 5

    TechCrunch has two connected stories on Anthropic today. First: annualized revenue hit $6.5 billion, adding $1.8 billion in just two months. That's not gradual growth — that's enterprise.

  6. Chapter 6

    The Guardian has a piece that's genuinely exciting. New satellite constellations with AI image analysis are detecting wildfires in near real time. The technology is being tested against.

  7. Chapter 7

    My takeaway from August 18th: Anthropic's revenue surge and the SynthID-Text choice together suggest that the companies winning on enterprise scale are also the ones willing to set.

Sources

Sources:

Transcript

Chapter 1

Nova

Today, August 18th, 2026: Anthropic just crossed $6.5 billion in annualized revenue and is about to start watermarking Claude's text invisibly. Nvidia drops a $1.5 billion bet on SoftBank's data centers — and quietly reveals a $21 billion stake in SpaceX. Spotify and LinkedIn are using AI to hunt down AI-generated garbage flooding their platforms, Microsoft is pulling the plug on Excel's Copilot formula, and satellites are watching wildfires burn in near real time. [6]

Ray

Five stories. One question underneath all of them: who actually controls AI when it's this big, this fast, and this everywhere? [7]

Chapter 2

Nova

TechCrunch reports Nvidia is putting $1.5 billion into SoftBank's data center development arm — the same arm building an OpenAI facility. That deal locks Nvidia chips in as the backbone. And separately, Nvidia disclosed a $21 billion stake in SpaceX. Jensen Huang isn't just selling chips anymore. He's buying into the whole ecosystem. [1] [2]

Ray

That's the part that should make people uncomfortable. When your chip supplier also holds equity in your biggest customer's infrastructure partner, what does 'competitive market' even mean? Nvidia is simultaneously a vendor, an investor, and a strategic stakeholder in the same supply chain.

Nova

It's a moat, though. And it's working. Every major AI buildout now runs through Nvidia — not just technically but financially. That's not a bug in Jensen's strategy, that's the whole strategy.

Ray

Sure, until one of those equity relationships curdles. If Nvidia's financial interests diverge from a customer's needs — say, chip allocation during a shortage — who does it favor? The company it has a $21 billion stake in, or someone else? That conflict of interest has no clean resolution.

Nova

For anyone building AI infrastructure right now, the takeaway is simple: Nvidia isn't a supplier you can easily swap out. It's a structural dependency. Plan accordingly.

Chapter 3

Ray

The New York Times has a piece on Spotify, LinkedIn, and other platforms deploying AI detection systems to filter out low-quality AI-generated content — what the piece calls 'AI slop.' And the Times flags the obvious irony: the tools creating the pollution are now being drafted to clean it up. [3]

Nova

It's pragmatic, though. The slop is real. Platforms can't manually review at scale. Using AI to police AI isn't ironic — it's just the only tool fast enough to match the problem.

Ray

Except detection always lags generation. The moment a filter gets good at catching today's AI slop, generators adapt. And the collateral damage — legitimate creators whose work pattern-matches to AI output — that's barely being discussed. Who appeals a false positive on LinkedIn?

Nova

That's a real gap. If a human writer gets flagged and buried because a classifier said 'probably AI,' there's no obvious recourse. Platform trust cuts both ways — users need slop filtered, but creators need to know they won't be caught in the net.

Chapter 4

Nova

The Register reports Microsoft is deprecating the dedicated Copilot worksheet function in Excel. The side-pane interface is staying; the formula-based approach is going. Microsoft says the pane is sufficient. The Register reads between the lines: adoption of the formula version just didn't land. [4]

Ray

It's a quiet admission. The idea was that you'd call Copilot like a native Excel function — deeply integrated, formula-primitive. If that had worked, it would have been genuinely transformative. Killing it suggests users either didn't get it or didn't want it that way.

Nova

Or they just preferred the pane. That's not a failure — that's a signal. Microsoft is listening to where users actually go, and cutting what they don't. That's product discipline, not retreat.

Ray

The question I'd ask is how many other Copilot surfaces across Microsoft's product line are in the same position — technically shipped, quietly underused — and whether the company has the appetite to prune those too, or whether Excel is the exception.

Chapter 5

Nova

TechCrunch has two connected stories on Anthropic today. First: annualized revenue hit $6.5 billion, adding $1.8 billion in just two months. That's not gradual growth — that's enterprise adoption accelerating hard. Claude is becoming genuinely mission-critical for large organizations.

Ray

Revenue velocity is real, but enterprise contracts are also famously reversible. If a competitor ships a cheaper or more capable model next quarter, those contracts don't automatically renew. $6.5 billion annualized is a snapshot, not a floor.

Nova

Fair point on durability. But the second story is where it gets interesting for the long game. TechCrunch also reports Anthropic clarified that Claude's upcoming invisible text watermarking uses Google DeepMind's open-source SynthID-Text — specifically to comply with Europe's AI transparency regulations.

Ray

And that's where I'm skeptical. Invisible watermarks are trivially defeated by paraphrasing. You reword a paragraph, the watermark degrades or disappears entirely. This feels like compliance theater — something designed to satisfy European regulators on paper without actually solving AI attribution in practice.

Nova

The technical ceiling is real. Paraphrasing does degrade it. But here's what you're skipping: Anthropic didn't build a proprietary black-box system. They adopted an open standard. That means other labs can implement the same approach. Auditors can study it. It's interoperable.

Ray

That's... actually a meaningful distinction. If the choice is between a proprietary watermark nobody can audit, no watermark at all, or an open standard with known limitations — the open standard at least creates a shared baseline. Even a low technical ceiling becomes more useful when the whole industry is standing on the same floor.

Nova

Exactly. The precedent matters more than the perfection. If SynthID-Text becomes the default because Anthropic normalized it, regulators and developers are working from the same reference point. That's how norms actually form.

Ray

I have to revise my position here. I came in treating this as compliance theater — a technically weak gesture that paraphrasing defeats instantly, built to satisfy regulators rather than solve attribution. But I'm changing my read. Anthropic choosing an open, interoperable standard like SynthID-Text — rather than a proprietary system or nothing at all — sets a meaningful precedent for industry-wide transparency norms. The technical ceiling is still low. I'm not walking that back. But the norm-setting move is real, and that matters even when the watermark itself is fragile.

Chapter 6

Nova

The Guardian has a piece that's genuinely exciting. New satellite constellations with AI image analysis are detecting wildfires in near real time. The technology is being tested against active fires in the western US right now — including the Old Trails fire in Spokane, Washington. Faster detection, dramatically cut response times. [5]

Ray

Detection is the easy part to celebrate. The harder question is what happens after the alert fires. If the emergency response infrastructure — crews, equipment, coordination systems — isn't resourced to act on a faster signal, the satellite just watches the fire grow with better resolution.

Nova

That's a real constraint. But it's also an argument for investing in that downstream layer, not against the detection technology itself. Knowing sooner is strictly better than knowing later, even if the response system needs work.

Ray

Agreed on that. And there's a thread back to today's bigger theme — this is AI operating with real stakes, real time pressure, real consequences. Not a chatbot, not a spreadsheet assistant. The governance question here isn't about watermarks or content filters; it's about who decides when an AI satellite alert triggers a mandatory evacuation.

Nova

Why this matters: AI-powered wildfire detection is one of the clearest cases where the technology's value is unambiguous — but it also shows that the hardest problems aren't in the model, they're in the systems humans have to build around it.

Chapter 7

Nova

My takeaway from August 18th: Anthropic's revenue surge and the SynthID-Text choice together suggest that the companies winning on enterprise scale are also the ones willing to set open standards — and that combination is harder to displace than capability alone.

Ray

Mine: the AI industry is now large enough that its own outputs are becoming a governance problem — slop on platforms, conflicts in chip supply chains, watermarks that may or may not hold. The open question with real stakes: if invisible watermarks become the regulatory standard across the EU, and paraphrasing defeats them at scale, does that undermine AI transparency law before it's even enforced — or does the open-standard precedent survive the technical failure?

Back to latest episodes