(Preview) SaaSmageddon and the Future, Microsoft After a Market Correction, Anthropic’s Super Bowl Lies

Summary

• A reader’s bear case cast Microsoft’s $350 billion market-value loss as the software sector’s warning: labs own the models, enjoy a claimed “70% lower marginal cost,” and can hijack incumbent distribution. Andrew Sharp largely agreed, but Ben Thompson challenged the premise that labs will be superior “on every vector”—enterprise software is often bought for controls, compliance, and accountability rather than AI experience.
• The durable SaaS moat may reside in the 2% of cases where probabilistic systems fail, not the 98% where an LLM produces a better interface. Thompson’s healthcare example was Epic: its hated forms encode HIPAA, drug interactions, regulation, and liability, making it “superior on a vector that everyone hates” but that drives the bottom line.
• Enterprise software institutionalizes repetitive processes because one mistake can erase years of small efficiency gains. Thompson’s calendar example showed why: scheduling by a guest’s time zone seems reasonable until he edits an event distractedly and misses the setting; a rigid process assumes “I am going to screw up” and prevents the costly exception.
• Microsoft’s product shortcomings are not new, which paradoxically leaves Thompson less bearish than the market correction implies. “Why do we think Microsoft was gonna be good at this when they’ve sucked at products forever?” he asked, placing the company in the same skepticism cycle previously faced by Google, Apple, and Meta.
• Defending existing software moats does not justify assuming nothing changes when code becomes dramatically cheaper to produce. Thompson compared that input shock with the internet eliminating newspaper distribution costs: an apparent expansion in addressable market ultimately removed local monopolies and exposed every publication to power-law competition.
• The newspaper analogy is imperfect because software moats are more layered, but Thompson considers it plausible that AI could eventually discern and handle those rules on the fly. User-generated content took roughly 30 years to become most people’s media consumption—perhaps “99%,” setting aside television to some extent; AI may take longer than current expectations, though fast-moving technologies can punish that confidence.

Deep dive

1. AI labs do not win on every enterprise vector

• Reader Rav’s indictment was specific: Microsoft had three years of access to OpenAI’s IP, yet ChatGPT beat Copilot in enterprise settings and GitHub Copilot was “stuck in 2024” behind Cursor, Windsurf, and Claude Code. Because labs own the costly models, he argued, applications cannot beat rivals with “a 70% lower marginal cost.”

• Sharp found himself nodding through nearly all of it, correcting only Rav’s metaphor: Microsoft would be the “canary in the coal mine,” not a red herring.

• Thompson’s pushback — worth keeping: superior at what? If the job is delivering the best AI experience, labs win “definitionally.” Enterprise applications, however, often persist precisely because user experience is not the buying criterion; they solve difficult, mostly invisible problems involving risk, controls, and edge cases.

2. The ugly 2% is enterprise software’s moat

• Thompson’s explanation for why bad applications survive: their visible ugliness reflects invisible requirements. A vendor understands the industry’s “muck,” encodes its controls, and lets a CIO feel confident that employees using Excel will not accidentally “sink the company.”

• Healthcare is the extreme specimen. Epic installations force doctors through “50 gazillion boxes and forms,” but those fields sit downstream of HIPAA, drug interactions, regulation, and massive liability. Thompson called a clean interface “definitionally impossible” when software must manage that many variables and identify responsibility when something goes wrong.

• He conceded the system may be net worse because doctors drown in paperwork instead of treating patients; he also described Epic as entrenched after solving requirements downstream from Obamacare. Yet this remains “the vector that everyone hates” and “the vector that actually drives the bottom line.”

3. Institutionalized process protects against one fat finger

• Thompson illustrated the moat with interview scheduling across time zones. A colleague used each guest’s local zone, which made conversational sense, but Thompson wanted his own zone encoded because he might edit an event while distracted and miss the setting: “We need to have a process that assumes I am going to screw up.”

• That tiny example scales into the “400 SaaS apps” inside a company. Buttons and forms are institutionalized processes: paying $50 per seat may look absurd when a task takes three minutes manually, but “one screw-up costs you a bunch of money.”

• LLMs create a “superior experience” because they are probabilistic, loose, and right most of the time. Embedded in that fluency is error; much traditional software exists specifically to eliminate its possibility. Saving a little repeatedly can be completely undone by the 2%.

4. Microsoft’s old weakness meets a genuinely new input shock

• Thompson’s blunt framing was not that Microsoft suddenly became bad at AI products: “Why do we think Microsoft was gonna be good at this when they’ve sucked at products forever?” Product excellence has not been its advantage for decades, which paradoxically supports treating this as another big-tech skepticism cycle, not necessarily a terminal verdict.

• Still, Thompson would not let the SaaS defense become complacency. When software becomes dramatically cheaper to produce, “you don’t get to say nothing’s going to change because software is really important.” A fundamental input changed, so market structure will change with it.

5. Zero-cost distribution shows how an advantage becomes exposure

• Before the internet, newspapers were effectively manufacturing and trucking businesses. Going online appeared to expand The Washington Post’s market from the D.C.-Maryland-Virginia region to the world; in reality, the expensive distribution system it viewed as a constraint had protected its geographic monopoly.

• Once distribution approached zero, every publication gained the same reach. Power laws followed: readers generally chose one insufficiently differentiated national subscription, and “everyone subscribes to The New York Times,” leaving the Post competing for the rest.

• Entry costs also collapsed, so the Post came to compete with Thompson and other publications for subscriptions, and with Facebook, TikTok, and YouTube for finite attention. Thompson noted that few people in 1993 anticipated user-generated content eventually accounting for perhaps “99%” of most media consumption, setting aside television to some extent; the transition took about 30 years.

• Reader Marshall’s pushback was that newspaper defensibility proved shallow, while software has “stacked and multifaceted” moats, including switching costs. Thompson agreed the analogy underplays those layers, then preserved the unresolved risk: why couldn’t AI eventually discern all those regulations and rules on the fly? He thinks that is plausible and may take longer than expected, while admitting fast-moving technologies repeatedly punish such confidence.