Anthropic’s listing raised $81.3 billion, priced above its own upsized range, and became the largest IPO in history [CNBC][Associated Press]. Many will take the record as validation for every company built on top of frontier AI models. The logic goes that Wall Street has decided AI infrastructure is worth a historic premium, and anyone one layer removed from that infrastructure should benefit from the halo.
The details complicate that story. Needham’s own bull case for the stock assumes revenue growth decelerating from 156% to 98% within a year [Needham & Company]. Morningstar’s analysts put fair value somewhere between $500 billion and $3 trillion, a range so wide that it says more about the unresolved question of what a frontier AI lab is worth than about Anthropic itself [Morningstar]. The same market that just handed Anthropic a record valuation can’t agree on that valuation within a factor of six.
That split matters more than the record. A company can produce the largest IPO ever and still leave the market divided on how to price the layer it occupies.
Southeast Asia was working on the same question at almost the same time. Just days before Anthropic’s listing, ASEAN’s 49th Summit in Manila closed with the signing of the Digital Economy Framework Agreement (DEFA), a legally binding regional pact governing cross-border data flows and source-code protection for AI systems, more and more of which run on frontier models like Anthropic’s [ASEAN Secretariat]. APEC, meeting in the same window, adopted its own AI Cooperation Initiative through 2030, the first AI framework with both the US and China as signatories [APEC]. Three different institutions were effectively negotiating one question at roughly the same time: as frontier AI gets priced, regulated, and distributed globally, who in the value chain gets to keep the value?
For Southeast Asia, the consensus answer has been repeated in analyst notes and conference panels for two years, often enough to become received wisdom. The region has no frontier labs, so the value will sit in its application layer. Build on top of Claude, GPT, and Gemini. Let Anthropic and OpenAI burn the capital while SEA captures the margin downstream.
That answer is half right, and as a strategy it is dangerously incomplete.
The Compression Moves Up the Stack
Anthropic’s own numbers argue against complacency. The company has committed roughly $900 billion to compute infrastructure, nearly $70 billion of it due within the next twelve months [The New York Times]. Goldman Sachs, the lead bank on the IPO, has privately forecast the company will burn through more than $65 billion in negative free cash flow in 2027, with operating losses peaking around $40 billion that year, followed by continued heavy spending into 2028 [Quartz][The New York Times]. None of that capital is flowing to an “application layer” in Jakarta, Ho Chi Minh City, or Manila. It’s going to Nvidia, AWS, and TSMC.
Meanwhile, the price of a given level of AI performance has fallen roughly 13-fold a year since 2023, according to Epoch AI’s latest analysis — a rate the research group found to be faster than any other transformative technology in history, including compute, DNA sequencing, and lithium-ion batteries [Epoch AI]. An earlier Epoch analysis found even sharper swings of 9x to 900x a year depending on the benchmark. However it’s measured, the direction is the same: capability that was state-of-the-art and expensive eighteen months ago is now routinely available for a fraction of a cent. The orchestration and application layer that Southeast Asia has quietly bet its AI strategy on is next in line for that compression. Anthropic is already moving up the stack into agents and managed workflows — Claude Code, Claude Agents billed by the session-hour on top of token fees — into the territory SEA’s AI companies have staked out as their own.
The useful question, then, is which companies in the application layer have built something that doesn’t get cheaper and more replaceable at the same rate as the models underneath them.
Nearly every SEA AI startup now calls itself “model-agnostic” in its deck. Under that test, the label stops being an automatic badge of safety. A router that swaps Claude for Gemini or DeepSeek by changing an environment variable is a feature any hyperscaler will bundle for free within two product cycles, and there is nothing in it to defend. A moat, if one exists, has to come from somewhere the frontier labs’ price wars can’t reach: proprietary workflow data, regulatory and compliance infrastructure, or deep localization that OpenAI or Anthropic will never find commercially worth building themselves.
Applying that test to four companies operating in this stack in Southeast Asia gives a clearer picture of where the value sits.
The Infrastructure Layer: Cloudsway
Cloudsway sits closest to the most exposed part of the stack. It provides GPU cloud access, model routing across providers, and managed services for enterprises deploying AI. That puts it in the “intelligent routing and model portability” category, which analysts increasingly flag as the first thing hyperscalers will absorb and give away as a free feature.
This doesn’t make Cloudsway’s position weak by default, but it does mean being precise about where its value comes from. The defensible part of the business is guaranteed, reliable compute access and managed infrastructure relationships in a region where GPU capacity is still scarce. The MSP relationship — someone an enterprise can call when something breaks — doesn’t commoditize the way an API call does. With compute becoming the binding constraint, as Anthropic’s own $900 billion infrastructure commitment shows, this layer’s value will concentrate in access and reliability rather than routing logic.
The Workflow Layer: fileAI
fileAI is a cleaner example of a moat that survives model commoditization. Any frontier model can already do rudimentary document extraction, so “we use AI to read documents” is not the pitch. Its moat is the enterprise-specific workflow built on top: document taxonomies, reconciliation logic, and compliance workflows for KYC and financial processing. These have been developed and refined in real client environments, at a scale the company reports as roughly 200 million files processed a year, for clients including Toshiba, KFC, and MS&AD.
That workflow doesn’t get cheaper when Anthropic cuts token prices. Falling inference costs should improve fileAI’s margins, since the underlying model call costs less while the workflow around it holds its value. The pure-routing version of Cloudsway’s business faces the opposite exposure. Side by side, the two show that the “application layer” is several layers, each with very different exposure to the same commoditization curve.
The Orchestration Layer: Diaflow
Diaflow is the hardest of the four to place, which makes it the most useful test case. Its architecture is explicitly model-agnostic, bundling OpenAI, Anthropic, Gemini, Llama, Mistral, and Cohere into a single orchestration layer that it markets as a way to avoid vendor lock-in. Taken alone, that would put it in the weak position described above: a router.
Diaflow’s actual bet is on governance. It offers SOC 2, HIPAA, and GDPR-compliant infrastructure, workflow logic, and agent orchestration that enterprise customers find hard to rip out. Switching models is easy; switching an orchestration system with compliance and process logic embedded in it is not. The test is whether a customer could replace Diaflow with a generic agent-builder tomorrow. If the governance and workflow layer is doing real work, they can’t — and Diaflow counts as infrastructure rather than a router.
The Localization Layer: WIZ.AI
WIZ.AI offers the sharpest example of a moat frontier labs structurally cannot compete away, because trying wouldn’t make economic sense for them. Customers in Southeast Asia speak Singlish, Bahasa, Tagalog, and Thai, and often switch between them within a single conversation. Anthropic and OpenAI aren’t optimizing for that. Next to English-language enterprise and developer markets, hyper-localized Southeast Asian voice AI is too small a market to justify frontier-lab attention.
WIZ.AI’s business fills that gap, and it has built the moat deeper than orchestration alone. Rather than relying solely on routed calls to frontier models, WIZ.AI has developed its own smaller, language-specific models tuned for the region’s linguistic range, including Singlish, Bahasa, Tagalog, and Thai. These models are built to run with far fewer parameters than a general-purpose frontier model, trading raw general capability for density of regional language intelligence, optimized specifically for the conversational patterns, code-switching, and accents of Southeast Asian speech. A smaller model trained tightly on this distribution can outperform a much larger general model on the narrow task that actually matters commercially: understanding a customer who moves between three languages in one sentence. It also means WIZ.AI’s unit economics in its core market aren’t hostage to frontier-lab pricing at all; the intelligence doing the heavy lifting is its own.
That combination compounds. Its localized conversational AI reportedly processes over 100 million calls a month across 300-plus enterprise clients, and the company recently closed a Series B led by SMBC. Its moat comes from data, language, and proprietary model architecture built specifically for this distribution, not from borrowed frontier capability. It is built through volume and specificity, compounds the longer it runs, and doesn’t erode when a frontier lab halves its API price, because much of what it runs was never renting that price in the first place. This version of the application layer was never at risk of commoditization, because general-purpose intelligence — the thing being commoditized — was never its product.
Where DEFA Fits
It would be easy to treat ASEAN’s Digital Economy Framework Agreement as incidental timing, a regional bureaucratic milestone that happened to land in the same news cycle as Anthropic’s IPO. That undersells it.
DEFA’s provisions on cross-border data flow governance, and on source-code non-disclosure as a market-entry condition, will make it easier for Anthropic, OpenAI, and Google to commercialize in Indonesia, Vietnam, or Thailand [ASEAN Secretariat]. They also structurally favor the governance-and-compliance layer that fileAI and Diaflow are building. Financial services, healthcare, and government are under growing pressure to adopt AI while meeting local data residency and compliance requirements. For them, companies that have already built SOC 2/HIPAA/GDPR-grade infrastructure and local workflow expertise become close to structurally necessary. A foreign frontier lab can sell an API into the region. It cannot, on its own, satisfy a local regulator that the resulting workflow is compliant, auditable, and governed. That takes the layer these companies occupy.
This dynamic sits inside a larger geopolitical frame. APEC’s AI Cooperation Initiative — the first AI framework to carry both US and Chinese signatures — signals that Southeast Asia is positioning itself as a convening ground between the world’s two AI superpowers, not merely a market they both sell into [APEC]. DEFA, in that light, is not just a compliance checklist; it is ASEAN asserting the terms on which frontier intelligence gets to operate inside its borders. For the application-layer companies built on local data, local workflows, and local regulatory relationships, that assertion is a tailwind, not a constraint.
So besides laying rails for frontier AI, DEFA helps build the moat for the companies sitting between frontier models and regulated local demand.
The MSME Gap Frontier Labs Won’t Close
There is a second, quieter signal from the same summit worth noting: ASEAN’s accompanying Leaders’ Declaration on AI-Powered MSME Growth, which frames affordability and access for small and medium enterprises as a regional AI priority. Anthropic’s commercial roadmap, by contrast, continues to skew toward enterprise and developer tooling — agentic coding products, large-contract enterprise deployments, the kind of customer that can absorb usage-based pricing at scale.
That gap between where frontier labs are building and where ASEAN’s stated policy priorities sit is exactly the space the application layer should be filling. WIZ.AI’s localized voice AI, already serving mid-market enterprise clients across the region and running on its own smaller language-specific models rather than solely on rented frontier capacity, is one version of closing that gap — and arguably the more durable one, since it isn’t priced off frontier-lab economics at all. fileAI’s document workflows, deployed against the unglamorous back-office reality of KYC and reconciliation, are another. Neither company is competing with Anthropic for frontier capability. Both are doing something frontier labs structurally will not prioritize: making AI usable for the businesses ASEAN’s own policy language says it wants to serve first.
The Real Takeaway
Anthropic’s IPO will be remembered as the moment public markets decided frontier AI deserved a record premium. For Southeast Asia, the more durable lesson sits one level down. Being “the application layer” was only ever a starting position. Companies that treated it as a finish line — pure routers and wrappers with nothing underneath but a model API — are exposed to the price and capability compression that Anthropic’s own numbers describe. Those that used it to build something that holds its value as the models get cheaper, whether a workflow, a compliance layer, a proprietary language model, or a trust relationship, are the ones positioned to compound.
Southeast Asia’s AI companies were never competing to be Anthropic, and the smarter ones never treated rented intelligence as a business model. As ASEAN builds the regulatory rails and APEC negotiates the geopolitical terms under which frontier AI operates in the region, the companies that will matter most are not the ones with the cleverest router. They are the ones that own the parts of the stack that keep their price when tokens lose theirs.
References
- CNBC — “Anthropic prices IPO at $83 a share, raising $81.3 billion in largest-ever listing”
- Associated Press — “Anthropic goes public in record-setting IPO”
- The New York Times — “Anthropic’s Compute Commitments and Cash Burn, Explained”
- Quartz — “Goldman Sachs Models Anthropic’s Path to Profitability”
- Morningstar — “Anthropic Fair Value Estimate: A Wide Range Reflects Deep Uncertainty”
- Needham & Company — Equity Research Note on Anthropic (via financial media coverage)
- Epoch AI — “Trends in AI Inference Cost and Capability” (2025 and 2026 analyses)
- ASEAN Secretariat — “ASEAN Digital Economy Framework Agreement (DEFA) Signed at 49th ASEAN Summit, Manila”
- APEC — “APEC AI Cooperation Initiative 2026–2030, Adopted at Gyeongju Leaders’ Meeting”
- Company disclosures and public materials — Cloudsway, fileAI, Diaflow, WIZ.AI
Paulo Joquiño is a writer and content producer for tech companies, and co-author of the book Navigating ASEANnovation. He is currently Editor of Insignia Business Review, the official publication of Insignia Ventures Partners, and senior content strategist for the venture capital firm, where he started right after graduation. As a university student, he took up multiple work opportunities in content and marketing for startups in Asia. These included interning as an associate at G3 Partners, a Seoul-based marketing agency for tech startups, running tech community engagements at coworking space and business community, ASPACE Philippines, and interning at workspace marketplace FlySpaces. He graduated with a BS Management Engineering at Ateneo de Manila University in 2019.