Martin Källström
knowledge / philosophy

Growth strategy & risk

Martin's growth philosophy is not a single doctrine but a lesson learned twice, in opposite directions. Twingly (2006) was built to survive on modest capital indefinitely; Narrative (2012–2016) was built to bet everything on speed, raising every six months and betting the company that being first would matter more than being careful ▶ 12:29 ▶ 4:15. Narrative went bankrupt; Twingly didn't. The thesis Martin draws from both experiences is that the worst outcome for a company isn't failure — it's treading water for years without growth or a decision ▶ 12:55 — but that going big also demands the discipline to notice when the market has stopped agreeing with you, which he admits he failed to do in time. He now applies both lessons well beyond hardware, using them to read strategic moves across the current AI industry.

Go big or go home vs. sustainable growth: the central bet

The choice of growth strategy is itself a strategic decision, made early and rarely revisited. With Twingly, Martin's strategy was to raise 10 million SEK in 2007 and make it last as long as possible ▶ 12:29 — slow, stable growth by design ▶ 12:19. With Narrative he deliberately chose the opposite: a 'go big or go home' strategy, betting heavily to find out fast whether the idea was right or wrong ▶ 11:47 Breakit ↗. Each Narrative funding round was planned to last exactly six months, forcing an immediate pursuit of the next round to keep momentum and market timing on the company's side ▶ 4:15.

Strategy coherence matters more than speed — choosing 'go big or go home' creates investor belief and momentum, but it also commits a company to a path that becomes extremely hard to reverse. Changing a company's core strategy mid-course is extremely difficult and requires great perseverance, which is part of why Martin kept pursuing Narrative's original strategy even as market signals deteriorated ▶ 26:12. Investor pressure was itself a force pushing toward unsustainable growth: Narrative's CTO later attributed the company's troubles directly to investors pushing a pace the business couldn't sustain TechCrunch ↗, and the reformed, post-bankruptcy Narrative deliberately prioritized slower, lower-risk growth instead TechCrunch ↗.

By contrast, Google's own 1999 Series A is the counter-example Martin now reaches for: rather than spending aggressively, Google saved most of that money in the bank ▶ 26:37 — proof that 'go big' and 'spend fast' aren't the same decision. Underneath both strategies sits a belief that entrepreneurial breakthroughs are ultimately decided by luck, and that the real founder skill lies in maximizing the number of chances you create for that luck to land — itself an argument for the aggressive path: more shots on goal, faster.

The danger of believing your own hype

The greatest danger for a successful company is believing its own hype. Positive press coverage can obscure a founder's awareness of broader market dynamics and external threats ▶ 18:43. Despite heavy public hype around Narrative, Martin and the team did not monitor the wearables market's decline closely enough ▶ 19:09. The signal, in hindsight, was unmissable: GoPro released a catastrophic earnings report the day before Narrative's own Series A meetings were scheduled to begin, in January 2016 ▶ 20:27. Martin now regrets not adjusting Narrative's aggressive growth strategy once investor sentiment toward wearables had clearly turned ▶ 25:45, and believes better monitoring of investor sentiment toward the category might have let the company pivot or moderate in time Breakit ↗.

Founders become blind to macro market shifts precisely because they're deeply focused on their own product — which is why Martin argues external monitoring of competitor and investor sentiment has to be a deliberate, ongoing practice, not an afterthought. He connects this to a structural problem in venture capital itself: VC incentives push startups to build hype around a vision before the product exists, creating a systemic cycle where even truthful founders end up over-promising ▶ 28:33.

He has since lived this lesson from the other side, too. The excitement of a San Francisco AI hackathon led him to abandon a voice-AI project of his own about two months from beta ▶ 12:37 ▶ 14:19 — a decision driven by OpenAI's advanced-voice demo announcement ▶ 13:27 that he now deeply regrets, since that OpenAI demo itself sat unshipped for months — hype rather than reality ▶ 14:19. The trip, he concludes, was simultaneously valuable for inspiration and counterproductive for the clarity and energy his own project needed ▶ 14:45.

Margins, runway, and timing the bet

Narrative's bankruptcy left Martin with a permanently changed filter for evaluating any new business idea: whether it can reach 30–40% product margins, checked before validating or launching anything, his own ventures included ▶ 28:38. A company cannot be sustainable long-term without margins in that range ▶ 29:03, and he now argues startups should sell at adequate margins much earlier than most do, building growth and resilience together rather than trading one for the other ▶ 29:03. He carefully weighs unit economics and margin potential before backing any new idea now, his own or someone else's ▶ 29:56.

Runway length is part of the same calculation. Narrative's earlier fundraising round provided capital for 24 months; Martin believes 36 would have been better Breakit ↗ — a direct, numeric version of the same lesson as Twingly's 'make 10 million SEK last as long as possible' approach. He was too aggressive with Narrative's growth strategy and, in hindsight, should have been more cautious Breakit ↗.

The practical version of this discipline for software today: start with commercial APIs for speed and production-readiness, and only move to self-hosting once you hit a specific limit — flexibility, cost, or performance ▶ 21:53. Use APIs for infrequent calls, self-host for always-on workloads ▶ 23:31 — the same instinct as choosing horizontal cloud scaling over buying servers: don't spend capital solving a scale problem you don't have yet.

Related: Narrative for the full company history.

Born global: international ambition from day one

Swedish startups fail to scale globally not from a lack of ideas, but from a lack of ambition to be global from day one — and going international from day one is no costlier than staying local: pitching US tech bloggers takes the same effort as pitching Swedish journalists ▶ 52:27. Both of Martin's companies were built 'born global' rather than local-first ▶ 6:24, though they pursued the strategy differently. Twingly focused on Sweden as its home market first, which made early PR traction and customer access easier but also capped early growth potential ▶ 6:50; it then expanded from European blogs into the US market only once established TheNextWeb ↗. Narrative went further, establishing its PR firm in the USA and targeting America as its home market from the outset ▶ 29:24.

A geographically remote base was no obstacle. Narrative was built out of Linköping, a small city in peripheral Sweden, and still achieved global reach through its Kickstarter campaign ▶ 4:45 — the internet, Martin argues, is precisely what makes an international brand buildable from anywhere ▶ 5:10.

Scaling architecture: horizontal over vertical

Global scalability must be baked into architecture from day one, not bolted on later once data is too large to restructure. Martin's clearest articulation of this comes from Twingly's own infrastructure story: cloud services alone don't guarantee scalability — the application architecture itself has to be designed from inception to exploit multiple servers or added capacity ▶ 11:05. Scalability, performance, and availability are three separate properties that must be optimized separately: a service can be scalable without being fast, and scalable without being highly available ▶ 8:58.

Horizontal scaling via hash sharding is economically rational for startups — it trades complexity for cost savings and lets a company scale in step with its cash flow rather than a large upfront capital outlay. Horizontal scaling is more attractive than vertical scaling because smaller companies can afford to start with modest resources and grow incrementally ▶ 14:36, and offloading infrastructure to third-party clouds turns a scaling problem into someone else's problem: Twingly could not have scaled to its traffic volume at all without cloud infrastructure ▶ 6:50, and never needed to buy additional servers or worry about capacity limits from the start ▶ 34:02. Twingly's own in-house indexing servers saw far more downtime than Amazon S3, which had only two outages between 2006 and 2010 ▶ 32:44 — though S3 has unpredictable download times and needs a CDN in front of it when performance matters, since it's a storage service, not a delivery one ▶ 35:43.

That architecture proved itself under real load: Twingly served the combined traffic of all 115 customers on the exact same technical solution it had deployed three years earlier for its first two customers, DN and SvD ▶ 6:24. Twingly's stable infrastructure proves that application architecture, not frequent rewrites, sustains growth at scale. Martin still turned down at least one high-visibility opportunity when the risk didn't pencil out: he declined Aftonbladet's request to put Twingly Live on their homepage during Melodifestivalen because building the necessary horizontal-scaling change in one week was too risky ▶ 42:36. Cloud hosting has real trade-offs too — it limits hardware experimentation, like testing SSDs against traditional drives ▶ 45:40, and cloud servers can crash and need rebooting, so applications must be designed for failure with storage kept separate ▶ 46:05.

Not every counter-example favors horizontal scaling: FriendFeed's success without it is Martin's reminder that vertical scaling — simply upgrading hardware — can be cost-optimal when commodity hardware is already powerful enough for the job, and that long-polling at scale (hundreds of thousands of concurrent open connections) needs a fundamentally different server architecture than either approach alone.

Idea generation and execution discipline

Ideas are cheap and interchangeable; what matters is execution, passion, and the disciplined work of turning possibility into reality. Martin's own practice reflects this literally: while working part-time at Twingly, he systematically developed 45–50 business ideas in a spreadsheet, comparing and categorizing them daily ▶ 36:17. His advice to other entrepreneurs is the same discipline at a slightly different dose: spend months generating around 40 ideas before committing to one, purely to gain perspective on your own true motivations ▶ 71:31, or — in a faster cadence — spend only about three days per idea, to keep iterating rather than getting impatiently attached to any single one ▶ 42:09.

He describes the selection process itself in violent, vivid terms: a 'Hunger Games of entrepreneurship' where competing ideas must fight until one victorious idea emerges strong enough to completely capture the founder's mind ▶ 1:51 — the winning idea should leave you standing 'on top of the hill, bloody and sweaty and heaving from the battle' before you ever start building it ▶ 15:28. Genuine belief in an idea has to be communicated with real conviction to make it real to anyone else ▶ 72:21, and once it wins, it should completely consume your mind — colonizing your perception and your conversation.

The fundamental problem-solving skill is asking 'who can solve this?' rather than 'how do I solve this?' — reframing constraints as a network problem rather than a personal-capability one ▶ 35:23. And don't work in secret: Martin advocates talking openly about startup ideas, arguing that secrecy kills companies by preventing the serendipitous connections that open discussion creates ▶ 48:46.

The right time to start a company is when you're young and your cost of living is lowest — delaying entrepreneurship for a 'normal job' trades away the exact window when the risk is cheapest. Martin urges students to use their education stipend as a first funding step ▶ 31:47, and suggests fresh engineering graduates negotiate four-day work weeks with employers specifically to reserve one day a week for their own startup ▶ 56:15. Underneath the advice is a personal urgency: losing his parents taught him that you don't live forever, and he can't wait until later to live his dream European CEO ↗.

Team composition is the other half of execution. Successful businesses require four core elements: a powerful idea, a capable team, market creation through a multi-channel launch, and personal conviction ▶ 21:54 — and a startup team needs a perceived division between technical and marketing expertise, even when individuals actually have both skill sets ▶ 23:11.

Hardware startups & crowdfunding: validate before you can't change course

Hardware startups must validate their product with customers before launching a major Kickstarter campaign, because after launch it becomes nearly impossible to change. Once a campaign is live, changing the product becomes extremely difficult, since backers inevitably hold conflicting preferences about which features matter ▶ 15:50. Thorough pre-launch preparation — pitching to at least 100+ potential customers — is what makes crowdfunding success possible at all ▶ 2:40, and hearing real objections in those conversations let Martin refine both the campaign video and landing page before backers ever encountered the same concerns ▶ 3:31.

Preparing for a Kickstarter campaign requires enormous, often underestimated effort — and success brings its own crisis. Narrative was unprepared for the overwhelming customer-support load that followed its Kickstarter's success, requiring round-the-clock response efforts ▶ 11:33. Communication is essential to maintaining community value and group morale when customer support becomes overwhelming ▶ 13:03. In hardware startups generally, most development and manufacturing activities take 1 to 1.5 years almost regardless of the specific task ▶ 19:33, and the underlying business passes through a rough succession from an R&D-driven company into a sales-oriented one, with target unit volumes scaling roughly tenfold each generation — 25,000 units for the first, 250,000 for the second, 2.5 million for the third ▶ 19:48.

Crowdfunding itself is more than a fundraising channel — it delivers market intelligence and validates product-market fit better than any survey, because it forces customers to put money behind their stated interest ▶ 6:07 ▶ 7:09. Martin treats it as one launch strategy among several — Kickstarter, traditional media, direct sales — to be deliberately chosen, not defaulted into ▶ 8:21, and stretch goals as a planning and market-testing tool rather than just a backer-engagement gimmick Kickstarter ↗. Even so, Memoto missed its original April shipping deadline and shipped seven months late, on November 1, 2013 Gizmodo ↗, despite the Kickstarter itself exceeding funding expectations by $500,000 Gizmodo ↗.

The get-to-market discipline Martin distilled from that experience: get the minimum viable product right Rude Baguette ↗, get it to market quickly and inexpensively, then iterate on real feedback rather than delaying launch to chase feature completeness Rude Baguette ↗ — he deliberately delayed promised features like GPS location browsing at Narrative's launch specifically to avoid scope creep SlashGear ↗.

Reading strategy in the current AI wave

Martin applies the same bet-the-company lens to companies he's only observing. Google failed to disrupt its own search business despite inventing the foundational AI technology — a textbook case of the innovator's dilemma: Google invented the transformer architecture five years before OpenAI, and despite the largest dataset and search traffic in the world, failed to capitalize on it ▶ 5:11. He sees the pattern repeating: existing companies generally won't lower prices when they capture AI efficiency gains, which is exactly the opening startups will exploit — Klarna replacing Salesforce and other large SaaS platforms by automating processes with AI instead of licensing more software is his go-to example ▶ 27:16, and he extends the same logic to Microsoft, Windows, and PCs as AI computer-use capability changes how people use operating systems at all ▶ 27:16.

Don't build the foundational layers of AI infrastructure; instead focus on what will be uniquely valuable after commoditization occurs. Martin's own strategic reasoning for time-to-market delays is that they can become an advantage rather than a cost: entrepreneurs should trust that technological progress in AI — lower costs, higher speed, more capacity — follows predictable patterns and will happen naturally ▶ 3:30, so it's wasted effort to optimize for a trend that will improve on its own ▶ 3:30; delayed time-to-market lets a founder capture better AI capability and lower cost by the time they actually launch ▶ 3:59. At Multiply, that logic became a deliberate planning exercise: an onsite specifically to map which AI capabilities were likely to simply be handed to them by the industry's progress, versus what the team needed to build themselves ▶ 7:39.

That patience has a hard limit, though — the iteration imperative: ship today rather than wait for better tools, because the rate of improvement means waiting is strategically pointless. Martin uses a space-travel analogy to make the case for starting projects now rather than waiting for future AI improvements, since better tools will always arrive later regardless ▶ 29:15, and stresses shipping and testing with real use cases immediately rather than delaying for improvements that haven't landed yet ▶ 30:03. The two positions aren't a contradiction: trust natural progress on the underlying technology, but don't let that trust become an excuse to stop shipping product.

Startups can compete in AI through pioneering focused technology, or through niche, relationship-driven applications that mainstream players won't bother building — and either way, entrepreneurs must genuinely innovate beyond thin API wrappers, because relying on shallow integration with someone else's model is not a defensible position on its own ▶ 27:48. Martin's clearest example of durable differentiation is Copy.ai: its edge came from UX design and prompt engineering as intellectual property, not the underlying AI technology ▶ 18:09, built by narrowly targeting marketers and breaking writing tasks into roughly 100 specific use cases rather than offering a general-purpose tool ▶ 26:18 — a company that reached $10 million in annual recurring revenue on UX innovation, not novel AI capability ▶ 17:59. By contrast, the speech-to-text and text-to-speech market was oversaturated with at least 25 competing startups per category the last time Martin checked ▶ 10:31 — the difference between a startup built on original ideas with long-term vision and a thin wrapper destined to be short-lived ▶ 12:49. find.com is his example of a product protected from commoditization by a genuine proprietary asset: indexed documentation from more than 200,000 open-source projects ▶ 14:05.

Adopt new protocols and technologies when organizational need is urgent, not out of curiosity or strategic speculation ▶ 29:29 — the same discipline, applied to technology adoption, as checking margins before chasing an idea. And OpenAI's own behavior is, in Martin's reading, a strategy worth studying on its own terms: he characterizes the company as the 'new vacuum cleaner of innovation,' rapidly copying into ChatGPT whatever a significant competitor builds — citing Perplexity's open-web publishing feature and Anthropic's Canvas/document model as things OpenAI absorbed ▶ 10:18.

See also: AI business & value and AI-native interfaces & products.

Worth remembering