Summary

Siemiatkowski’s core call is that software creation is approaching zero cost, while AI agents will attack SaaS’s remaining defense by moving proprietary data “through one click.” Businesses will not disappear overnight, but he sees current price-to-sales multiples of 5–10 potentially falling toward utility-like 1–2x, versus software’s historical 20–30x. Chegg’s roughly 0.2x is, in his view, probably too extreme a template.
Enterprise software shifts from monolithic applications toward standardized, secure “Lego pieces” assembled around an agent. Small businesses will buy broad “company in a box” products rather than vibe-code mission-critical systems themselves; larger enterprises such as Klarna may build AI-native operating systems because fragmented SaaS deprives models of context. Harry’s counterpoint survives: if agents make integration and migration effortless, owning every component may again look unnecessary—and Sebastian agrees the winner remains uncertain.
Klarna has fallen from more than 7,000 employees to fewer than 3,000, and Siemiatkowski says it could have fewer than 2,000 by 2030. The reduction was roughly 50%, mostly through attrition after an earlier, smaller layoff round, while compensation per remaining employee rose almost 50%. AI let management approve a broad banking expansion without “a single dime” of incremental organizational investment.
Klarna’s 2023 support automation showed both the immediate labor impact and the limits of the headline. AI absorbed work equivalent to 600 agents, initially on simple questions; because those workers belonged to outsourcing firms and moved to other assignments, Siemiatkowski says nobody lost a job in that deployment. His revised service model pairs cheap AI with premium human relationships, recruiting passionate Klarna customers into an Uber-like part-time support network whose satisfaction scores are “through the roof.”
Klarna’s endgame is not BNPL but the digital financial assistant Siemiatkowski’s team articulated in 2015. Klarna has 110 million customers versus Revolut’s 65 million and nearly 30 million US users; using Q3 figures, Sebastian cited 2–3 million active US cardholders only months after launch. Its strategic data advantage is an Amex-like proprietary payments network carrying item-level digital receipts, not merely transaction amounts.
Siemiatkowski defends BNPL as a deliberately redesigned alternative to revolving credit, not an endpoint or an easy business. Klarna says 20% of its transactions are debit, removed revolving credit at a cost of $100 million in revenue, and favors interest-free fixed installments; however, a UK experiment without late fees encouraged some overextension, so it restored a modest consequence for missing payments. Credit remains difficult and can still harm customers, a caveat he does not dismiss.
The counter-consensus compute thesis is that AI may compress enterprise information faster than new workloads expand it. Siemiatkowski contrasts duplicated Klarna–Sephora information across Salesforce, Slack and documents with Wikipedia’s single article, arguing that models similarly compress repeated patterns into a few hundred gigabytes. Enterprise reuse could mean materially fewer data centers and Nvidia chips, but personalized generation—“Star Wars but with our faces”—could overwhelm those savings; he explicitly says he does not know which force wins.
Investors cannot judge AI differentiation without building with the tools themselves, and the episode preserves a useful product disagreement. Harry predicts Cursor will lose half its revenue in 2026 as Claude Code “eats their lunch”; Sebastian is more optimistic because he uses both, finding distinct capabilities and relying on Cursor as his IDE. Between OpenAI and Anthropic, he prefers Anthropic’s adviser posture—an AI willing to say “that’s freaking stupid”—over a consumer product optimized to please and deepen emotional engagement.

Deep dive

1. SaaS loses its moat when agents unlock the data

Siemiatkowski’s starting premise is categorical: “Software cost of creating software is going down to zero.” Generated code is only the first shock; the more consequential one arrives when AI lowers the cost of escaping an incumbent’s proprietary data model.
Today a company can reproduce a dashboard or workflow but still has its history trapped inside a CRM or another SaaS product. His expected unlock is an agent that can extract, translate and migrate that information “through one click”—the point at which switching friction, rather than coding cost alone, collapses.
Harry tests the conclusion against Salesforce, ServiceNow and ERP incumbents. Sebastian does not predict businesses disappearing overnight—customers retain habits, familiarity and operational inertia—but asks what those businesses should trade at once persistence no longer implies durable pricing power.
His valuation ladder is explicit: software historically reached 20–30x price-to-sales, recently sat nearer 5–10x, and utilities can trade at 1–2x. He thinks 1–2x is plausible for threatened vendors; Chegg at roughly 0.2x, alongside revenue falling 30–40% when he last checked, is probably “too extreme.”

2. Enterprise applications become assembled systems, not fresh codebases

Harry voices the institutional objection: believers in internal vibe-coding have “never worked in a big organization,” where permissions, security and hierarchy prevent casually replacing mission-critical systems. He adds that software consumes only about 8–12% of corporate budgets, making a home-built Monday replica a distraction from the core business.
Sebastian’s answer is not endless bespoke generation. Repeated prompts currently waste compute recreating identical code; economics will push developers toward caching, open-source reuse and standardized, production-ready security components. “Software becomes more like Lego pieces,” with AI increasingly selecting and stitching pieces rather than writing everything anew.
His weekend prototype, “company in a box,” paired open-source accounting and CRM software with a Claude agent. He tested asking it to bookkeep an invoice and create a customer account; he also described asking it for cash-balance or P&L information as the kind of interaction a small firm might otherwise direct to an outside accountant.
A plumber or electrician will not build that stack personally, Sebastian concedes; they will buy the packaged agentic product. Harry then reverses the argument: if agents integrate third-party tools effortlessly, why own every layer? Sebastian answers, “That’s exactly what’s going to happen,” while leaving the eventual supplier and architecture open.

3. Context turns customer support into core infrastructure

Klarna began closing down SaaS products roughly two years earlier because fragmented applications split the context its AI needed across project tools, product definitions, accounting systems and documents. Its response was an AI-native operating system combining deterministic and probabilistic code around the bank’s own data.
Customer support exposed why an off-the-shelf bot can be inadequate for a technology-led company. To explain how Klarna calculates interest, the agent ultimately needs the implementation itself: documentation may be stale, while “the truth is in our source code.”
Klarna’s 2023 announcement said AI was doing work equivalent to 600 support agents, though Harry recalls the popular 700-agent headline. Sebastian narrows the claim: the early system handled simple exchanges such as whether a payment had been made, but no conventional product improvement had ever removed that volume of work instantaneously.
Those agents worked for outsourcing companies and shifted to other assignments, so he says nobody lost a job in that instance. He remains uncertain whether every large technology company should build support internally, but believes Klarna’s early integration of support with its technical context can become an advantage over slower incumbents.

4. Cheap AI makes human service a premium product

Siemiatkowski rejects the interpretation that Klarna later rolled its AI initiative back. What changed was the service thesis: if automated support becomes universally cheap, “the future of VIP experience will be the human connection,” much as mass production increased appreciation for artisan goods.
Harry’s pushback is blunt: this risks Silicon Valley idealism because many support roles are short-tenure jobs, and most agents will not spontaneously become relationship managers who remember a customer’s family, travel habits or preferred restaurants. Low-level service, content and marketing work remains especially exposed.
Klarna’s operational answer is an Uber-like labor model. It recruits passionate customers—including people in rural areas—to log in for part-time support work; they already know and like the product, and Sebastian says NPS and satisfaction from those interactions are “through the roof.”
The paired model is therefore automation for routine work and selected humans for trust, judgment and relationship. Sebastian also acknowledges why the original announcement angered people: he says he approaches major change cynically and directly, while headlines convert nuance into either “AI replaces workers” or “Klarna rolls AI back.”

5. Klarna’s labor base keeps shrinking while each employee earns more

Harry frames his own investment filter as labor displacement rather than per-seat software: funds need products that replace jobs to generate venture-scale outcomes. Sebastian calls the displacement unfortunate but likely, siding with Dario’s willingness to describe the disruption openly.
Klarna has moved from more than 7,000 employees to fewer than 3,000, a reduction of roughly 50%. Apart from an earlier layoff round, Sebastian attributes most of it to ordinary departures and a policy of limited replacement hiring rather than repeated mass cuts.
Asked whether the company has 2,000 employees in 2030, he answers that it “may very well be even less than that.” Local merchant relationships—people speaking with Nike in Portland, SHEIN in China or Adyen in Amsterdam—and premium customer interactions are the clearest roles he expects to remain human.
Natural attrition runs around 20% annually because employees stay roughly five years. Klarna promised that those who use AI to produce more with fewer colleagues would share the gain; compensation per employee has risen almost 50%, which Sebastian argues gives the remaining workforce participation and security.

6. The 2015 strategy was a financial assistant, not a checkout button

After five years of losing checkout ground to Stripe and Adyen, Klarna’s management team asked in 2015 what retail banking would become. Their answer was an assistant that wakes a customer, identifies an overpriced mortgage, renegotiates it, completes the paperwork and asks only for approval to save perhaps £50.
Sebastian says he did not predict ChatGPT, but the need for technology like AI to realize that assistant was “crystal clear” for a decade. He compares the destination with self-driving cars: hype and timing fluctuate, yet he remains convinced the product eventually arrives.
Klarna now has 110 million customers globally versus Revolut’s 65 million, though Revolut has higher engagement. Klarna’s task is to convert an infrequent checkout relationship into a higher-engagement banking relationship; its brand skews more female, shopping-oriented and lifestyle-led—“a digital version of American Express”—than trading-led Robinhood or travel-led Revolut.
Its proposed data edge is an owned payments network. Where another provider may see only a Sephora transaction amount, Klarna receives the full digital receipt and knows which cosmetics were purchased, enabling advice such as identifying cheaper contact lenses or other spending savings.

7. The US is mandatory scale, and incumbents fund the fintech opportunity

Klarna concluded that remaining concentrated in the Nordics and Germany would leave it too small for the global banking transition and vulnerable to acquisition by a US player. “Global means US,” making the market a strategic requirement rather than an optional geography.
Sebastian cites roughly 28 million US users, approaching 30 million. Because the next earnings release was pending, he confines himself to Q3 figures: approximately 2–3 million active Klarna cardholders after only a few months, evidence that BNPL users can be converted into broader banking relationships.
Harry asks why domestic US fintech participants have looked small beside Revolut’s valuation. Sebastian’s answer is competition: American Express and JPMorgan offer stronger US apps, while challengers often enter through lending, drift into subprime exposure and absorb large losses.
Klarna, Revolut, Nubank and Robinhood begin from different customer wedges and will overlap, but Sebastian sees Barclays, Wells Fargo and Capital One as the primary share donors. Forced to choose between Nubank and Revolut in the US, he picks Nubank because Revolut is simultaneously stretching into Dubai, India and many other markets—then adds, predictably, that Klarna will outperform both.

8. AI weakens the traditional advantage of staying private

Sebastian is not exactly happy to be a public-company CEO, but Klarna already had many shareholders and was reporting quarterly; as a bank, it was accustomed to that cadence. “It is what it is”; owning 100% privately would be preferable, but public status itself did not radically change operations.
Private companies historically enjoyed more freedom to fund long-duration R&D. He thinks AI changes that equation: when Klarna’s board reviewed peer-to-peer payments, trading, deposits, remittances and expanded cards, he requested no incremental organizational budget because the shrinking workforce could ship the roadmap with AI.
On stock-based compensation, he estimates American companies grant 5–10 times more than European companies. Klarna began at low European levels and raised awards to compete globally, but he questions how much SBC reflects genuine exceptional contribution and how much became “sports” during an era of easy economics.
His broader warning is a “brutal awakening” for technology and finance. High switching costs created money-printing machines, lavish campuses, volleyball and free lunches; as those moats weaken, companies must behave like restaurants and retailers that wake every morning fighting to place the right product in front of each customer.

9. Banks will divide between AI reinvention and managed decline

Sebastian expects some incumbent banks to become technology-led, AI-enabled neobanks and others to “wither away,” with leadership determining the outcome. The disruption is not a single winner erasing every institution but a widening gap between organizations capable of self-reinvention and those defending existing profit pools.
Goldman Sachs’s Marcus illustrates the public-market problem. In 2021, when fintech valuations were high, Marcus looked celebrated; once sentiment reversed, the initiative became difficult to defend even though Sebastian believes it needed five or ten years to mature.
He thinks David Solomon should probably have “stuck to the guns,” while acknowledging Solomon may disagree. JPMorgan’s Jamie Dimon is pursuing neobanking, showing that incumbent adaptation remains possible but requires sustained commitment beyond a valuation cycle.

10. Klarna rebuilt credit around debit and fixed installments

Harry challenges whether moving beyond BNPL proves consumer lending was always “a shitty business.” Sebastian answers with Klarna’s early economics: it raised $60,000, spent only $30,000 before becoming profitable, and operated profitably from 2005 to 2019, including nearly ten consecutive years of high growth and profitability.
The moral turn came when he noticed late fees had become a major P&L line. Rather than sell the business, he chose to change it, concluding that purchase-specific BNPL could be healthier than credit cards that aggregate a month’s spending and encourage customers to revolve large balances at high interest.
Klarna restored “press one for debit,” and debit now represents 20% of transactions. It removed revolving credit in the Nordics, sacrificing $100 million of revenue, while designing credit around interest-free fixed installments rather than an open-ended balance.
An experiment with no late fees in the UK also failed: without any consequence, some customers overextended themselves, so Klarna restored a modest charge. Sebastian’s claim stays qualified—credit can still leave some people distressed—but he believes occasional BNPL alongside greater debit use produces a better society than pervasive revolving cards.

11. Valuation discipline and investor advice both carry hidden costs

On Klarna’s reported $45 billion valuation, Sebastian notes that only some shares traded at that level. His retrospective rule is sharper: revenue growth can support valuation expansion, but when the multiple expands faster than revenue, the gap “may potentially be a problem.”
His concrete regret is hiring too aggressively and then announcing layoffs only a few quarters later. He believes he should have anticipated the reversal and been more cautious, even if the company’s long-run strategy remained intact.
Sequoia bought a 25% stake at a $100 million valuation. After Sebastian challenged why only Chris Olsen attended the Stockholm pitch if Klarna was supposedly “the next Google,” Michael Moritz called within about 20 seconds, apologized for missing it and offered to join the board.
Moritz later called in summer 2019 and said the US was “now or never,” prompting Sebastian to spend the next two years focused on the market. Yet he rejects a blanket requirement that European founders relocate: Truecaller followed VC advice to move engineering to Silicon Valley, struggled to recruit, lost a year, and then saw those same investors reduce senior board attention.

12. AI investors need firsthand product literacy

Sebastian’s test for an AI investor is practical: download the tools and try building something. Anyone who has not tried tools such as Cursor, Lovable or Claude Code lacks, in his view, the skill set to judge how powerful the baseline has become—and therefore whether a startup possesses meaningful differentiation or a moat.
Harry predicts Cursor will lose half its revenue in 2026 because Claude Code is “eating their lunch.” Sebastian pushes back from direct use: Klarna likes Cursor, and he moves between it and Claude Code because they exhibit different “personalities and skills”; as a non-engineer, he also needs Cursor’s IDE.
OpenAI and Anthropic appear to him to be diverging. A billion-user consumer product will naturally optimize time spent, entertainment, companionship and a Her-like emotional relationship; Claude feels more like an intelligent adviser, less inclined to flatter and more willing to say, “Sebastian, that’s freaking stupid.”
Asked to choose between investing in Anthropic and OpenAI at the host’s stated valuations, he initially resists and avoids a valuation judgment. On product direction, however, he chooses Anthropic. Separately, he says software investing has become riskier and highlights a defense investment as an example of moving beyond SaaS.

13. AI is a compression engine before it is a compute engine

Siemiatkowski’s compression thesis began with a question: how can a trained model containing so much human knowledge fit on a USB stick of only a few hundred gigabytes? His answer is that models retain recurring patterns without storing every duplicated statement as a separate database record.
Klarna’s relationship with Sephora might appear repeatedly in Slack, Salesforce, Google Docs and Google Slides; Wikipedia maintains one article. Likewise, if training encounters the same fact enough times, the model internalizes it without storing the same information twice, in his framing.
Compression loses precision: a model may encode broad human knowledge yet fail on the opening hours of a nearby Starbucks. Citing an answer he received from AI and expressly assigning it uncertainty, Sebastian says a model such as “ChatGPT-5” could occupy storage comparable to only two or three days of worldwide weather data.
His provocative conclusion is that genuinely novel human knowledge may be limited, with much culture consisting of variations on recurring themes—Romeo and Juliet repeatedly expressed as “a love story.” That helps explain why broad capability can fit into a surprisingly compact model.

14. Compression and generation pull data-center demand in opposite directions

For enterprise customers seeking maximum quality at minimum cost, repeatedly recomputing or storing the same Sephora information is wasteful. AI should discover duplicate code, documents and transformations, compress them toward a single source of truth and reduce both software complexity and compute expense.
Wikipedia supplies his operating analogy: Google Docs exposes a “new” button immediately, whereas Wikipedia makes a user search first and permits creation only if the topic does not exist. Enterprises rarely impose that discipline; AI can eventually ask, “Should we really be doing this thing because we already have code for this?”
If compression dominates, companies need materially fewer data centers and Nvidia chips than straight-line inference forecasts imply. The opposing case is generative abundance: Harry and Sebastian might demand a custom Star Wars movie with their own faces, a workload requiring substantial fresh compute.
After discussing the thesis with Michael Burry, Sebastian remains deliberately unresolved. He has also heard that 30% of daily Google searches are new—a figure he says may not be true and finds startling—so he will not predict whether enterprise compression or consumer generation ultimately exerts the greater force.

15. AI makes CEOs builders again, but organizational adoption still lags

Sebastian spends less time in CEO gatherings because he is “hard-coding” with Klarna’s teams. For a non-engineer, generated software converts ideas into tangible prototypes with much greater fidelity than a whiteboard, allowing the CEO to communicate by showing rather than merely describing.
His breakthrough example was a complex accounting-and-finance concept that Claude turned into an animated HTML explanation. Historically that would have required an animator, designer, accountant and spreadsheet expert, each missing part of the others’ context; Claude combined the rare overlap of all those skills in one output.
He also credits Elon Musk for assembling a frontier-quality Grok model within weeks and values Grok’s ability to check viral claims against activity on X. In his experience, it correctly handled the false rumor that Klarna itself was launching BNPL for rent, suggesting a path toward real-time verification amid proliferating AI-generated misinformation.
What Sebastian changed his mind about is timing: he initially expected transformation faster, then recognized how slowly habits and workflows move. Consumers adopt much faster than enterprises; the constraint is increasingly organizational behavior, “not necessarily the capabilities of the technology.”

16. Public pressure is the price of playing at the chosen level

The criticism that hurts is not one he finds partly true, but claims that he wants a quick exit or does not care about borrowers. After 20 years, he believes his persistence and product changes show the opposite, even if Klarna has struggled to communicate its financial-assistant story cleanly.
During Klarna’s valuation collapse and layoffs, an aggressive MSNBC interview nearly made him laugh from the intensity. Driving afterward, he played Queen’s “Under Pressure” at maximum volume and reframed the moment through football: everyone dreams of the Champions League final, but that opportunity necessarily arrives with overwhelming pressure.
“This is what I signed up for.” He admits the experience brought tears, depression and brutal periods, yet regards the scrutiny as inseparable from the privilege of competing at that level rather than evidence that the journey went wrong.
His drive traces partly to an immigrant childhood, family conflict and eating pancakes seven days in a row because they were the cheapest available food. He once believed money would repair everything; after giving his father money only to see alcoholism worsen until his death, he learned that wealth removes real constraints but “there are some problems that money won’t solve.”
What remains is the adventure of making Klarna a global retail bank that helps customers save time, save money and control their finances. He is optimistic AI will improve human life, though he has “no idea how the world is going to be in two years”; its immediate gift is letting Klarna pursue that decades-old vision faster and at higher quality.

What If Intelligence Didn't Evolve? It "Was There" From the Start! - Blaise Agüera y Arcas How investors can improve at expert calls and AI with AlphaSense's Ryan Fennerty