Two Sources of Irreplaceability: From DeepSeek and Kimi to Southern European Tourism

Deep Analysis · Technology & Tourism · October 28, 2026

From DeepSeek and Kimi to Southern European Tourism

Dr. Tong Yin, Chief Research Officer, InsightBridge Global and InsightBridge Global Lab.

The Question

A familiar claim in competitive strategy holds that a firm must either be the lowest-cost producer or the most capable one. The generative AI industry offers a test case that is still unfolding. Two Chinese companies, DeepSeek and Moonshot AI (maker of Kimi), have taken different positions in the same field. The first is built on cost efficiency, the second on the capability frontier. Both have won recognition from users and investors.

This essay asks why both positions can hold, where the irreplaceability of each comes from, and what limits each faces. It then turns to a field that looks unrelated, tourism in southern Europe, to show how the same logic works with physical resources.

Path One: A Position Built on Cost Efficiency

In January 2025, DeepSeek released R1, an open-source reasoning model comparable to OpenAI's o1. It rose to the top of the free charts on Apple's US App Store, and Nvidia's shares closed down nearly 17% that day (CNBC). The company said it trained the model on a modest budget, relying mainly on engineering and algorithmic efficiency (Fortune).

In 2026, DeepSeek open-sourced a preview of V4, built around low-cost million-token context (DeepSeek). V4-Pro was reported to cost about $0.87 per million output tokens, compared with about $50 for Anthropic's Fable model (Fortune).

The irreplaceability of this position comes from its cost structure. For developers and companies with heavy, price-sensitive workloads, comparable usable capability is offered at a small fraction of the price. A competitor would need a matching breakthrough in engineering efficiency to catch up; a price cut alone would not suffice. Open-source release widens the user base further and makes the model a default foundation for many applications.

Path Two: A Position Built on the Capability Frontier

Moonshot AI chose another route. Kimi K3, an open-weight model released in July 2026, has 2.8 trillion total parameters, activates about 104 billion per token and supports a million-token context (NVIDIA; Kimi API Platform). K3.1 Pro, released in late October, scored 55 on the Artificial Analysis Intelligence Index, ranking fourth behind models including Anthropic's Opus 5.5 (58) and Sonnet 5.5 (56) (Caixin).

Pricing reflects this positioning. Kimi's top international membership tier costs $199 a month (Kimi Help Center), expensive by the standard of Chinese models (Fortune). Even so, overseas platforms have chosen to integrate it. Perplexity added Kimi K3 to its products and hosts it on US servers (Perplexity), and Nvidia's model service lists it as well (NVIDIA).

In the author's observation, service elements such as access speed and API stability were not fully mature in Kimi's early overseas expansion, yet major platforms still chose to integrate it. When tasks are complex and errors are costly, buyers compare capability first, and service inconvenience comes second. A platform's willingness to bear the cost of local deployment is itself a price placed on that scarcity.

Figure 1: The cost curve and the capability frontier

Why the Two Positions Can Coexist

Capital markets have valued both paths highly. Reuters reported that Moonshot was valued at about $50 billion in its latest round and had confidentially filed for a Hong Kong listing, while DeepSeek was valued at about $74 billion (Reuters). Moonshot is reported to be targeting a Hong Kong listing in the first quarter of 2027 (Euronews).

Figure 2: Capital-market valuations of the two paths

The two positions can coexist because they serve different needs. Large-scale, standardized, price-sensitive workloads care about unit cost. Complex reasoning, long-document work and specialized professional tasks care about reliable results. In the first market, the barrier to substitution is the cost curve; in the second, it is the capability frontier. They are not the winner and loser of one contest but two distinct points of scarcity within one industry.

The two paths share one feature: they spread mainly through product performance rather than advertising. R1 spread because of test results after its open release, and K3 was integrated by platforms because of its performance in benchmarks and real tasks. This suggests a proposition: in markets where users can verify results directly, product performance is the most effective form of promotion; in markets with little substantive difference, where distinction rests mainly on marketing, the situation differs. The proposition needs testing with more industry data.

Talent as the Deeper Scarce Resource

Cross-border competition for top research talent (AI-generated illustration)

Behind both positions lies a concentration of top researchers. Moonshot founder Yang Zhilin was a co-first author of Transformer-XL and XLNet, two papers on long-sequence modeling (arXiv; arXiv). The company was founded in 2023; Alibaba led a $1 billion round in February 2024, and Tencent later joined (Wikipedia). Alibaba and Tencent each develop their own models yet still invest in outside teams, which can be read as option-like positions on frontier capability.

The US market shows the same pricing logic. WIRED reported that Meta offered some OpenAI researchers packages of up to $300 million over four years; Meta disputed the figures (Business Insider). Whatever the exact numbers, top researchers are priced as a scarce resource in both countries.

Limits and Risks

The strategy of irreplaceability does not apply everywhere. It works best where the cost of error to the customer is high, as in frontier AI, quantum computing, and professional services such as medicine and law. The larger the potential loss, the more customers will pay for the most reliable capability. In mass consumer goods, where substitutes are plentiful and errors cheap, the marginal return on technical excellence is limited, and competition turns more on cost, distribution and brand. A further hypothesis awaits testing: in large institutions, decision makers are often not spending their own money and may weigh a supplier's brand and compliance credentials more heavily, while owner-managed small and mid-sized firms focus more on verifiable results.

The risks also deserve attention. First, the source of a capability position must withstand scrutiny. In October 2026, the director of the White House Office of Science and Technology Policy alleged that Moonshot developed K3 by large-scale distillation of Anthropic's models (Washington Examiner). The allegation has not been independently verified, but it shows that if a capability is seen as resting on others' work, both its durability and its compliance standing come into question. Second, wide adoption brings scrutiny. In February 2025, Australia banned DeepSeek on government devices, citing security risks, after Italy had taken similar steps (Reuters). Third, the industry is still early, and technical approaches change month by month, so any lead is time-bound. Dependence on core teams and changes in governance and incentives after listing will also affect whether a lead can last.

Southern European Tourism: Physical Resources That Cannot Be Copied

Southern Europe's heritage, cities and coastline (AI-generated illustration)

The same logic appears in a very different field. Italy, Spain and Portugal have differing macroeconomic records. Italy's real GDP grew 0.5% in 2025, below the euro-area average of 1.3% (Istat). Spain grew faster, with the OECD projecting 2.9% for 2025, yet its unemployment rate stood at 10.5% in September 2025, the highest among OECD members (OECD). Portugal grew 1.9% in 2025 (IMF).

Their appeal to visitors has nonetheless remained steady. Italy has more World Heritage sites than any other country, 62 as of September 2026 (Guinness World Records). In 2025, Italian accommodation establishments registered 185 million arrivals, including 104 million foreign arrivals, up 8.7% (ANSA). These figures count accommodation registrations, so a visitor who stays in several places is counted more than once.

Price is another side of the appeal. With the EU average set at 100, the 2024 price index for restaurants and hotels was 76 in Portugal, 84 in Spain, 107 in Italy, 110 in France and 112 in Germany (Eurostat, reproduced by the Statistical Office of Montenegro). The Iberian Peninsula is clearly below the EU average; Italy is slightly above it but below France and Germany. Historic sites, cityscapes, food culture, coastline and relatively acceptable prices together form an overall experience that no single alternative destination can easily reproduce.

Figure 3: Southern Europe macro performance and tourism price levels

Assessments of service are mixed. The Bank of Italy's survey of international tourism shows foreign visitors rating most aspects of Italy between 8 and 9 out of 10, with prices rated lowest (Banca d'Italia). In 2019, about 90% of foreign respondents said their trip exceeded expectations (Banca d'Italia). The author's own field observation is that efficiency and responsiveness in some service areas still fall short; the survey data indicate that such shortfalls have not markedly lowered visitors' overall ratings or their willingness to come.

Venice is a more direct example. In 2025, Venice applied a booking fee to day-trippers on 54 days, and 723,497 payments were made. On the busiest day, May 2, 24,951 paying visitors arrived, more than half the city's resident population (Euronews). Demand on peak days stayed high even with booking and payment added, which suggests that for a destination that cannot be copied, demand is not very sensitive to price or convenience. The fee pilot is itself the city's response to the limits of carrying capacity.

The Common Logic

Placing AI models beside southern European tourism reveals three shared features of the strategy of irreplaceability.

First, the core resource determines the position. For Kimi it is the capability frontier, for DeepSeek the cost curve, for Italy its historic sites and cities. As long as the core resource cannot be copied, buyers keep returning.

Second, service gaps are tolerated, but not indefinitely. Kimi's early service limitations did not stop platforms from integrating it, and shortfalls in some Italian services have not stopped visitors from coming. Service is a complementary asset, however, and it affects length of stay, spending per visitor, repeat purchases and return visits. The core resource buys time; improving service determines whether that time becomes a lasting advantage.

Third, irreplaceability has limits. Technical leadership is constrained by time, compliance and the movement of talent; tourism resources are constrained by carrying capacity. Recognizing these limits matters as much as building an irreplaceable position.

Conclusion

The divergence between DeepSeek and Kimi shows that irreplaceability has more than one source. Extreme cost efficiency and an extreme capability frontier can both become positions that are hard to replace, provided they correspond to real demand. Southern European tourism shows that when the core resource itself cannot be copied, macroeconomic swings and service shortfalls are not enough to change the direction of demand in the short run. For firms and regions alike, the strategic question is not how to be best at everything, but how to identify the one point where they cannot be replaced, and to keep investing in it within its limits.