Five waves · Focus on now
The Agent Wave
Is Still Unfolding
Keep the five-wave map, but focus on the opportunity in front of us. Agents can build software and carry out work across applications. Better reliability, memory, tool use and coordination can expand what they handle. These waves overlap; science and robotics do not mean the agent wave is over.
Our thesis: agents are eating into software workflows, not just adding another chat box. The opportunity is to own the work from request to result — replacing some apps, connecting others, and building what does not yet exist.
Workflow references (not a forecast of total SaaS replacement): Anthropic · Microsoft Copilot Studio
Editorial analysis · Updated October 1, 2026. Dates mark broad waves, not exact starting lines. Linked sources support specific examples; the economic judgments are ours.
Wave 1 · 2016 —
Compute & Memory
Intelligence needs a physical substrate. Large-scale AI systems depend on compute, memory, networking and power working together. The economic question is where supply is hardest to expand — and how long that bottleneck lets a supplier retain pricing power.
The bottleneck gets paid. Until it moves.
AI is sold as intelligence, but delivered through physical infrastructure. NVIDIA’s Blackwell platform combines compute, high-bandwidth memory and interconnects. Our reading: the valuable unit is not an isolated chip specification but a working system delivered at scale. A faster processor does little for a customer who cannot power it, connect it or keep it busy.
Source: NVIDIA · Blackwell architectureCheaper intelligence does not automatically mean less hardware.
There are two forces to separate. Efficiency reduces the compute needed for a given task. Lower cost can also make more tasks worth automating. Agents add another variable: one request may trigger many model calls, tests and retries. The net result depends on whether useful demand grows faster than efficiency improves. Neither endless shortages nor collapsing demand follows automatically.
Memory is not a consolation prize.
Calling memory a helper confuses where a component sits with how much value it captures. Scarcity, qualification, yields and the cost of switching suppliers matter more than proximity to the chatbot. Korea should defend that industrial position while testing new chips against a harder question: does a customer deploy them repeatedly at a competitive total cost, including software migration and support? National origin is not a substitute for that answer.
- What to watch
- Repeat orders, utilization and delivered cost per useful task. Announced capacity is not the same as profitable demand.
- What could change this view
- If supply expands faster than paid usage, scarcity pricing can disappear while AI itself keeps growing.
Players & positions
Different positions in the same system: NVIDIA in accelerated computing, TSMC in manufacturing, SK Hynix and Samsung in memory.
Korea's position · Defend the bottleneck; prove the next one
Memory is a concrete industrial position, not merely a supporting role. The next question is how well that position survives changes in architecture and supply. For domestic accelerators such as Rebellions and FuriosaAI, distinguish technical promise from repeat customer deployments.
Wave 2 · 2020 —
Frontier Models
The model is the capability engine. But capability, distribution and profitability are different competitions. Closed services and open-weight alternatives put pressure on each other; the strategic question is which advantages survive the next model release, not who tops a benchmark this week.
A better model is not automatically a better business.
A benchmark lead creates an opening. Durable returns require something more: users who stay, developers who build around the platform, or an advantage in serving cost. Buyers care about quality on their own tasks, latency, reliability and price together. A slightly weaker model can be the better choice for a narrow workflow; a more expensive model can be cheaper overall if it avoids retries and review.
Open weights move the bill. They do not erase it.
Access to weights can give a buyer more control over deployment and customization, subject to the model’s license. It does not provide free serving, monitoring or security. The strategic effect is an outside option: buyers need not accept every closed-provider price or restriction. But owning the deployment also means owning the operational burden. Compare cost per accepted result, not a free download against an API invoice.
Sovereignty needs a purchase specification.
For Korea, the useful question is not whether a model is domestic. It is which dependency it removes. Offline operation? Control over sensitive data? A Korean-language task that alternatives fail? A credible sovereign program names the requirement, tests alternatives and accounts for continuing compute and talent costs. Research capacity can be worth funding without pretending every research model is already a competitive commercial platform.
- What to watch
- Retention, paid production usage and cost per accepted result on a fixed task set. A leaderboard is a capability signal, not a market-share table.
- What could change this view
- If model advantages remain large and hard to substitute, the model provider can retain more value than the applications built on top.
Players & positions
Compare model providers such as Anthropic, OpenAI and Google DeepMind alongside open-weight families such as DeepSeek and Qwen. These are different deployment choices, not a single permanent ranking.
Korea's position · Name the dependency worth removing
HyperCLOVA X, EXAONE, Solar and A.X should be judged against explicit customer requirements and alternatives. Sovereign capacity may have strategic value; commercial competitiveness still needs evidence. Neither a domestic label nor a declaration that Korea can never compete is a useful evaluation.
Wave 3 · 2024 — expanding
Agentic Labor
Agents are becoming a way to build software and get work done across it. They can use tools, coordinate steps and generate custom code, putting pressure on standalone SaaS features and the need to work through each app by hand. This wave is not finished: reliable long-running tasks, memory, integrations, permissions and evaluation leave substantial room to improve. Some software will be replaced; other software will become the data, services and systems of record agents depend on.
The interface is vulnerable. The ledger is harder to replace.
Separate three things a SaaS product sells: an interface, a way to execute work, and a trusted record of what happened. An agent can bypass the first and compete for the second while depending on the third. Asking an agent to update a customer record does not eliminate the CRM database, access rules or audit history. The exposed product is one whose main value is making a person click through a simple process. A system that owns essential records and controls can become more valuable even as its screens matter less.
Cheap code weakens the feature moat. It does not make ownership free.
If a team can generate its own dashboard or approval tool, a vendor cannot rely on that feature being expensive to reproduce. But the replacement still needs authentication, backups, upgrades and an owner when something breaks. The build-versus-buy line moves; it does not vanish. The best candidates for custom software are bounded workflows with clear inputs, tests and manageable consequences. Business-critical systems need a lifecycle budget, not just a successful afternoon demo.
The market can grow beyond today’s software budgets.
The upside is not limited to taking revenue from existing SaaS. There are processes handled with spreadsheets, email and manual follow-up because custom software was never worth commissioning. Lower construction and operating costs could make those processes economical to automate. That is new demand, not merely substitution. The constraint shifts toward discovering the real process, resolving exceptions and getting people to trust the result. A small workflow used every day can matter more than a broad assistant nobody depends on.
Outcome pricing is not the same as outcome alignment.
When agents do the work, charging only for human seats becomes a less natural fit. Usage and outcome fees offer alternatives, but the definition matters. Intercom’s Fin documentation includes configured handoffs among billable outcomes; an outcome does not always mean a problem solved without a person. Our test for any offer: what counts as success, who pays for retries, and how much human cleanup remains? A low headline fee can hide an expensive process. Measure accepted work after review, not activity generated by the agent.
Source: Intercom · Fin outcome definitionsReliability is the product, not the final polish.
A model can perform each step impressively and still lose the thread across a long job. Anthropic’s long-running-agent work describes explicit progress records, incremental execution and testing between sessions. This is a useful distinction: a capable model is not yet a dependable worker. Someone must define completion, preserve state, detect failures and resume safely. More agents do not automatically solve this; they can add handoffs, conflicting edits and review load. Parallelize independent work, not unresolved dependencies.
Source: Anthropic · Long-running agent harnessesAuthority is a separate axis from intelligence.
Illustrative workflow: an agent reads a request, prepares a quote, updates a draft CRM entry and proposes an invoice. Reading and drafting can run with limited permissions. Sending the quote or issuing the invoice crosses a different boundary: a customer receives a commitment. Require approval there, log the action and design retries so they cannot create duplicates. Sandboxes contain execution; they do not make a misleading email trustworthy. Faster work comes from clear boundaries, not from granting every tool unrestricted access.
A benchmark horizon is not a shift you can leave unattended.
METR measures task difficulty using the time a human expert would need, at a specified success probability. Its time horizon is not how long an agent can safely run alone, nor a direct measure of jobs replaceable. For a business, the useful evaluation is a representative set of its own tasks, including missing data and awkward exceptions. Track completion, correction time and failure severity together. A rising success rate is encouraging; an unobserved failure that sends the wrong payment is still unacceptable.
Source: METR · Time-horizon methodology and limitationsKorea’s opening is specific work, not another generic assistant.
A plausible starting point is a workflow that crosses Korean documents, local business systems and rules a generic assistant does not already understand. The valuable asset is not translation. It is authorized access, tested procedures, exception history and a customer willing to renew. Those advantages take operational work to accumulate. They must also survive an incumbent adding an agent. Ask why the customer would keep this product if the underlying model became twice as capable tomorrow.
- What to watch
- Completed workflows per reviewer-hour, cost after retries, customer renewals and the share of work requiring rescue. Compare against the old process, not against doing nothing.
- What could change this view
- If review and integration costs stay high, agents may remain copilots in many workflows. If incumbent systems bundle good agents, independent vendors must offer more than access to a model.
Players & positions
Coding agents are one entry point. Enterprise agents and workflow platforms extend the same pattern into business operations.
Korea's position · Build around real workflows
The practical opening is work that needs doing now: connect business systems, encode domain knowledge, and make agents reliable enough to use in production. Korean-language workflows and local integrations can be starting points, but defensibility must come from customer adoption, data access and measurable results — not nationality.
Wave 4 · 2026 →
AI for Science
AI can extend the agentic pattern into research: assembling evidence, proposing hypotheses and helping run experiments. Scientific progress still depends on reproducibility, specialist judgment and physical validation. This is an overlapping direction of development, not a reason to treat the software and agent wave as finished.
Hypotheses get cheaper. Proof does not follow automatically.
Google DeepMind’s Co-Scientist work pairs generated hypotheses with experimental verification. That distinction is the whole business problem. Producing more plausible ideas can move the bottleneck into experiment selection, lab throughput and replication. A research assistant that speeds literature review is useful, but it is not the same achievement as discovering a treatment that works. Judge each product at the stage it actually reaches.
Source: Google DeepMind · Co-Scientist researchThe experiment loop may be worth more than the prediction.
Our thesis is that a stronger position comes from repeatedly proposing, testing and learning from experiments the wider market cannot observe. Failed experiments matter too: they narrow the search. But this only becomes an advantage if the company has rights to use the data and a way to turn learning into paid results. For Korean bio-AI, ask who owns the evidence and downstream economics. A partnership announcement alone answers neither question.
- What to watch
- Prospective validation, independent replication and paid use beyond a pilot. Keep research productivity, experimental success and commercial approval separate.
- What could change this view
- If experiments remain the dominant cost and generated ideas rarely survive them, better models may improve research productivity without transforming development economics.
Players & positions
Separate evidence assistants such as OpenEvidence, hypothesis-generation research such as Co-Scientist, and drug-discovery platforms. They solve different parts of the research workflow.
Korea's position · Follow the validation and the rights
Standigm, Galux and Oncocross provide starting points for examining Korean bio-AI. Look past the AI label: which predictions were tested, which results were independently reproduced, and who owns the resulting data and commercial rights? Scientific promise and a repeatable business are separate milestones.
Wave 5 · Longer horizon
Embodiment
Robotics brings AI into physical tasks. Unlike a software rollout, deployment also requires hardware, safety checks, maintenance and workable unit economics. It is a longer-term opportunity to watch alongside software agents, not a fixed-date successor that makes today's agent work less important.
Sell a completed shift, not an impressive minute.
The relevant unit is reliable work over time. A robot that handles a demonstration still has to cope with interruptions, worn parts and environments that change. Count installation, supervision, downtime, maintenance and recovery in the cost. A narrow machine doing one job economically can be more commercially important than a general-purpose robot with a larger repertoire on video. Embodiment is an operations business as well as a model problem.
Manufacturing access is a starting point, not a verdict.
Korea can use real production environments to test whether a system saves money and survives daily use. The stronger signal is a second site buying after the first has operated, not a memorandum promising a future rollout. Hardware, control software and service providers may capture different parts of the value. There is no need to declare a national winner now, or to wait for this wave before improving the software workflows surrounding every factory.
- What to watch
- Repeat deployments, intervention-free operating time and fully loaded cost per completed task. Distinguish orders, deliveries and sustained use.
- What could change this view
- Hardware and service costs can absorb the gains from better AI. A capable robot is not automatically a cost-effective worker.
Players & positions
Examples to watch: Figure, Physical Intelligence, Tesla Optimus, Unitree and UBTech. Compare deployed work and reliability, not demonstration videos alone.
Korea's position · Potential, not a guaranteed win
Korea's manufacturing base offers useful capabilities in integration, components and production. Those capabilities do not guarantee an AI robotics winner. Demand, repeat deployments and service economics still need to be demonstrated; businesses need not wait for robots to benefit from agentic software today.
The practical conclusion
Own the result, not just the AI label.
For workers, start with a recurring task whose output you can check. For builders, find a workflow customers will pay to stop managing themselves. For executives and policymakers, fund measured deployment and useful capacity, not demo counts. Across all five waves, ask the same question: when intelligence gets cheaper, what scarce asset, responsibility or customer relationship does this company still own?
03 · From the big picture to the players