I’m not normally fond of year-end technology retrospectives, but 2025 was indeed a year of quantum leaps in the art of the possible and it has filled us all with measured optimism paired with some healthy and well-earned skepticism where AI is concerned. When I put architecture in perspective, I’m inclined to take a longer view of automation in all its variations over a decade. That’s why 2025 feels more like a footnote in a long series of events culminating in the perfect storm of opportunities we’ve been contemplating for some time now.

The composable infrastructure revolution

We’ve been moving toward self-aware, composable infrastructure in architecture for a while now and infrastructure-as-code was merely the first major inflection.

Let’s be honest, the old way of building IT infrastructure is breaking down. As an enterprise architect, the vicious cycle is very familiar. Tying agentic architecture demand-patterns to legacy infrastructure without careful consideration is fraught with peril. The old pattern is really predictable now: You provision systems, maintain them reactively and eventually retire them. Rinse and repeat.

That model is now officially unsustainable in the age of AI. What’s taking its place? Composable and intelligent infrastructure that can proactively self-assemble, reconfigure and optimize on the fly to match what the business needs.

For IT leaders, this shift from rigid systems to modular, agent-driven infrastructure is both a breakthrough opportunity and a serious transformation challenge. And the numbers back this up: the global composable infrastructure market sits at USD $8.3 billion in 2025 and is projected to grow at 24.9% annually through 2032.

What’s driving this hyper-accelerated growth? Geopolitical disruptions, supply chain chaos and AI advances are reshaping how and where companies operate. Business environments are being driven by reactive and dynamic agentic experiences, transactions and digital partnerships everywhere, all the time. Static infrastructure just can’t deliver that kind of flexibility based on marketing exercises that describe solution offerings as “on-demand,” “utility-based,” “adaptive” and “composable.” These are little more than half-truths.

A 2025 Forrester study commissioned by Microsoft found that 84% of IT leaders want solutions that consolidate edge and cloud operations across systems, sites and teams. As an architect in the consumer goods space, I found that our IT team would produce endless slide decks about composable enterprises ad nauseam, but infrastructure-as-code was the level of actual capability for some time.

Leaders wanted composable architecture that can pull together diverse components without hyperextended interoperability efforts. IBM’s research reinforces this, showing that companies with modular architectures are more agile, more resilient and faster to market — while also reducing the technical debt that slows everyone down.

The problem has been one of capacity and fitness for purpose. Legacy infrastructure and the underlying systems of record simply weren’t designed with agentic AI patterns in mind. My conversations with pan-industry architecture colleagues reflect the same crisis of expectation and resilience around agentic architectures.

Consider McKinsey’s 2025 AI survey that demonstrated 88% of organizations now use AI regularly in at least one business function and 62% are experimenting with AI agents. But most are stuck in pilot mode because their infrastructure can’t scale AI across the business.

If there are any winners in this race, they’ve broken apart their monolithic systems into modular pieces that AI agents can orchestrate based on what’s actually happening in real time.

AI agents: The new orchestration layer

So, what’s driving this shift? Agentic AI — systems that understand business context, figure out optimal configurations and execute complex workflows by pulling together infrastructure components on demand. This isn’t just standard automation following rigid, brittle scripts. Agents reason about what to assemble, how to configure it and when to reconfigure as conditions change.

The adoption curve is steep. BCG and MIT Sloan Management Review found that 35% of organizations already use agentic AI, with another 44% planning to jump in soon. The World Economic Forum reports 82% of executives plan to adopt AI agents within three years. McKinsey’s abovementioned State of AI research further highlights agentic AI as an emerging focus area for enterprise investment and describes AI agents as systems that can plan, take actions and orchestrate multi-step workflows with less human intervention than traditional automation.

As McKinsey puts it: “We’re entering an era where enterprise productivity is no longer just accelerated by AI — it’s orchestrated by it.” That’s a fundamental change in how infrastructure works.

IBM is betting big on this future, stating that “the future of IT operations is autonomous, policy-driven and hybrid by design.” They’re building environments where AI agents can orchestrate everything — public cloud, private infrastructure, on-premises systems, edge deployments — assembling optimal configurations for specific workloads and contexts. The scope of automation ranges from helpful recommendations to closed-loop fixes to fully autonomous optimization.

What composable architecture actually looks like

I recall no shortage of Lego-induced architecture references to composability over the last decade. Sadly, we conflated them with domain services and not how business capabilities and automation could and should inform how the Legos are pieced together to solve problems. Traditional infrastructure comes as tightly integrated stacks — hard to decompose, inflexible and reactive. The new composable model flips this, offering modular building blocks that agents can intelligently assemble and reassemble dynamically based on what’s needed right now.

Composability demands modularity and responsive automation

The foundation is extreme modularity — breaking monolithic systems into discrete, independently deployable pieces with clean interfaces. Composable infrastructure lets you dynamically assemble and disassemble resources based on application demands, optimizing how pooled resources get allocated and improving overall efficiency.

This goes far beyond physical infrastructure to include services, data pipelines, security policies and workflows. When everything is modular and API-accessible, agents can compose complex solutions from simple building blocks and adapt in real time.

Bringing cloud and edge together

Enterprise organizations are no longer treating cloud and edge as separate worlds requiring manual integration. The new approach treats all infrastructure — from hyperscale data centers to network edge — as a unified resource pool that agents can compose into optimal configurations.

McKinsey identifies edge-cloud convergence as essential for agentic AI: “Agents need real-time data access and low-latency environments. Combining edge compute (for inference and responsiveness) with cloud-scale training and storage is essential.” They further highlight how Hewlett Packard Enterprise (HPE) expanded its GreenLake platform in late 2024 with composable infrastructure hardware for hybrid and AI-driven workloads — modular servers and storage that let enterprises dynamically allocate resources based on real-time demand.

Agents running the show

Even IBM with its storied fixed-infrastructure history is all-in on agentic AI infrastructure capabilities — including agents and Model Context Protocol (MCP) servers — across its portfolio, making infrastructure components discoverable and composable by AI agents. These agents don’t just watch the infrastructure state; they actively orchestrate resources across enterprise data and applications, creating optimal configurations for specific workloads.

Management interfaces across IBM cloud, storage, power and Z platforms are becoming MCP-compatible services — turning infrastructure into building blocks that agents can reason about and orchestrate. Vendor-native agentic management solutions introduced similar AI-driven orchestration enhancements in 2024, letting large enterprises dynamically allocate resources across compute, storage and networking.

Self-aware and self-correcting infrastructure

Instead of manually configuring every component, composable architectures enable intent-based interfaces. You specify business objectives — support 10,000 concurrent users with sub-100ms latency at 99.99% availability — and agents figure out the infrastructure composition to make it happen.

Emerging intelligent infrastructure player Quali describes this as “infrastructure that understands itself” — systems where agentic AI doesn’t just demand infrastructure that keeps up, but infrastructure built from composable components that agents can understand and orchestrate.

Getting scale and flexibility in real time

Traditional infrastructure forces a choice: optimize for scale or build for adaptability. As architects, there are clear opposing trade-offs we must navigate successfully: Scale relative to adaptability, investment versus sustaining operations, tight oversight versus autonomy and process refactoring versus process reinvention.

Composable architectures solve this by delivering both. The dual nature of agentic AI — part tool, part human-like — doesn’t fit traditional management frameworks. People are flexible but don’t scale. Tools scale but can’t adapt. Agentic AI on composable infrastructure gives you scalable adaptability — handling massive workloads while continuously reconfiguring for changing contexts.

Self-composability and evolved governance

Agent-orchestrated infrastructure demands governance that balances autonomy with control. The earlier-mentioned MIT Sloan Management Review and BCG study found that most agentic AI leaders anticipate significant changes to governance and decision rights as they adopt agentic AI. They recommend creating governance hubs with enterprise-wide guardrails and dynamic decision rights rather than approving individual AI decisions one by one.

The answer lies in policy-based composition, defining constraints that bound agent decisions without prescribing exact configurations. Within those boundaries, agents compose and recompose infrastructure autonomously.

When AI agents continuously compose and recompose resources, you need governance frameworks that look nothing like traditional change management. A model registry that includes MCP connects different large language models while implementing guardrails for analytics, security, privacy and compliance. This treats AI as an agent whose decisions must be understood, managed and learned from — not as an infallible tool.

Making it happen in 2026

What should IT leaders do? Here are the most critical moves from my perspective.

Redesign work around agents first. Use agentic AI’s capacity to implement scalability and adapt broadly within parameterized governance automation rather than automating isolated tasks. Almost two-thirds of agentic AI leaders expect operating model changes. Build workflows that shift smoothly between efficiency and problem-solving modes.

Rethink roles for human-agent collaboration. Agents are an architect’s new partner. Reposition your role as an architect in the enterprise to adopt and embrace portfolios of AI agents to coordinate workflows, and traditional management layers change. Expect fewer middle management layers, with managers evolving to orchestrate hybrid human-AI teams. Consider dual career paths for generalist orchestrators and AI-augmented specialists.

Keep investments tied to value. Agentic AI leaders anchor investments to value — whether efficiency, innovation, revenue growth or some combination. Agentic systems are evolving from finite function agents to multi-agent collaborators, from narrow to broadly orchestrated tasks and other ecosystems and agents, and from operational to strategic human-mediated partnership.

The bottom line

The companies that will win in the next decade will recognize composability as the foundation of adaptive infrastructure. When every part of the technology stack becomes a modular building block and intelligent agents compose those blocks into optimal configurations based on real-time context, infrastructure becomes a competitive advantage instead of a constraint.

Organizations that understand agentic AI’s dual nature and align their processes, governance, talent and investments accordingly will realize its full business value. My architect’s perspective is that agentic AI will challenge established management approaches and, yes, even convince many of its ability to defy gravity. But with the right strategy and execution, it won’t just offer empty promises — it will deliver results. Further, our grounded expectations around the capacity of aging infrastructure and legacy demand patterns must guide us in ensuring we make intelligent decisions.

The question isn’t whether to embrace composable, agent-orchestrated infrastructure. It’s how fast you can decompose monolithic systems, build orchestration capabilities and establish the governance to make it work.

This article was made possible by our partnership with the IASA Chief Architect Forum. The CAF’s purpose is to test, challenge and support the art and science of Business Technology Architecture and its evolution over time, as well as grow the influence and leadership of chief architects both inside and outside the profession. The CAF is a leadership community of the IASA, the leading non-profit professional association for business technology architects. 

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

IDC의 최신 ‘전 세계 분기별 개인용 컴퓨팅 기기 추적 보고서’에 따르면 2025년 4분기 글로벌 PC 출하량은 전년 동기 대비 9.6% 증가했다. 다만 메모리 부족 사태가 기업과 소비자 시장 모두에 영향을 미칠 것으로 예상되면서, 출하량 증가세가 지속될 가능성은 크지 않은 것으로 분석됐다.

IDC는 메모리 수급 상황이 빠르게 변화하고 있다는 점을 고려할 때 12개월 뒤 PC 시장이 크게 달라질 것으로 내다봤다. IDC는 “이번 실적에는 윈도우 10 지원 종료로 인한 업그레이드 수요, 연초 관세에 대한 우려로 벤더가 당초 계획보다 많은 재고를 앞당겨 확보한 점이 영향을 미쳤다”라고 설명했다.

IDC 리서치 매니저 지테시 우브라니는 “메모리 부족이 산업 전반에 영향을 미치고 있으며, 그 여파가 향후 2년간 시장의 역학 구조를 재편할 가능성이 크다”라고 전망했다.

IDC는 지난달 공개한 보고서에서 “글로벌 반도체 생태계가 전례 없는 메모리 칩 부족 상황에 직면해 있으며, 이로 인한 파급 효과가 기기 제조사와 최종 사용자에게까지 확산돼 2027년까지 이어질 수 있다. AI 데이터센터에서 발생하는 수요가 공급을 계속해서 웃돌면서 디램(DRAM) 가격이 크게 상승했고, 이로 인해 수급 불균형이 심화되고 있다”라고 진단했다.

올해 PC 가격은 높은 수준 유지

IDC 연구진은 메모리 시장이 “수요가 공급을 크게 웃도는 전례 없는 변곡점에 놓여 있다”라며, “장기간 호황과 불황이 반복돼 온 산업이지만, 이번에는 양상이 다르다. AI 인프라와 워크로드가 빠르게 확대되면서 메모리 생태계 전반에 상당한 압박을 가하고 있다”라고 분석했다.

램 가격 급등은 기업용 PC 시장에도 당분간 영향을 미칠 가능성이 크다. 인포테크 리서치 그룹의 수석 자문 이사 테오 안토니아디스는 2026년은 물론 2027년까지도 가격이 높은 수준으로 유지될 것으로 내다봤다. 그는 “이로 인해 PC 가격은 최대 20%까지 오를 수 있으며, PC와 스마트폰 시장에서 명확한 승자와 패자가 갈릴 것”이라고 설명했다.

안토니아디스는 “장기 공급 계약과 같은 헤징 전략을 갖추고, 출하 물량이 많으며 충분한 현금 보유고를 확보한 벤더가 유리한 위치에 설 것”이라고 설명했다. 반면 “가격 경쟁력을 핵심으로 삼아 낮은 마진에 의존하거나 규모가 작은 업체는 이 기간 동안 어려움을 겪을 가능성이 크다”라고 전망했다.

대기업이 여전히 노트북은 3년, 데스크톱은 5년 주기로 교체하고 있는지, 아니면 사용 기간을 더 늘리고 있는지를 묻는 질문에 안토니아디스는 “아직 현실을 부정하고 있는 상태”라고 진단했다. 그는 “2026년 예산이 지난 분기인 2025년 4분기에 확정됐기 때문에, 아직 이 문제를 논의하는 고객을 거의 보지 못했다”라고 설명했다.

다만 그는 실제 교체 견적을 받는 과정에서 가격 변화를 체감하게 되면 일부 전략 수정과 구매 축소가 나타날 가능성이 있다고 내다봤다. 그는 “가격이 내려가지 않을 경우 2027년에는 저성능 PC를 활용한 VDI(Virtual Desktop Infrastructure)가 다시 부상하고, 노후 PC로 인한 생산성 저하를 상쇄하기 위해 클라우드 기반 솔루션을 다시 한번 확대하려는 움직임이 나타날 수 있다”라고 전망했다.

IDC 월드와이드 모바일 디바이스 트래커의 리서치 부사장 장 필리프 부샤르는 “일부 제조사가 이미 발표한 시스템 가격 인상 외에도, 메모리 재고를 유지하기 위해 평균적인 PC 메모리 사양이 낮아질 수도 있다. 앞으로 1년이 매우 변동성이 큰 시기가 될 것”이라고 내다봤다.

IDC는 지난달 보고서에서 메모리 부족 사태가 AI PC를 중심으로 한 업계의 성장 구조를 흔들 수 있다고 지적했다. IDC는 NPU를 탑재한 모든 PC를 AI PC로 정의하고 있는데, 이러한 기기는 본질적으로 더 많은 램을 요구하는 경향이 있다. 실제로 마이크로소프트의 코파일럿+ PC는 최소 16GB 램을 필요로 한다.

연구진은 소규모 언어 모델과 대규모 언어 모델이 점차 기기 내부에 탑재되면서 메모리의 중요성이 더욱 커지고 있다고 분석했다. 연구진은 “이에 따라 고급형 시스템 상당수가 32GB 이상으로 전환되는 흐름이 나타나고 있지만, 업계가 램 증설의 필요성을 인식하는 시점에 오히려 비용 부담이 지나치게 커졌다는 점이 문제다. 설령 공급을 확보하더라도 추가 비용이 과도해진 상황이다. 그 결과 가격 인상이나 마진 축소, 혹은 신규 시스템에서 램 용량을 낮추는 선택이 불가피해질 수 있다”라고 내다봤다.

우브라니는 지난 1~2년간 메모리 부족 문제가 없었음에도 기업용 AI PC 수요가 다소 주춤한 흐름을 보여왔다고 설명했다. 그는 “클라우드 기반 AI 옵션의 완성도가 매우 높아지면서, 많은 이용자가 클라우드 AI와 온디바이스 AI의 차이를 체감하지 못하고 있는 점이 주요 배경”이라고 말했다.

그는 “이번 메모리 공급난의 영향으로 벤더가 하이브리드 AI를 본격적으로 내세우기 시작할 수도 있다. 메모리 부족으로 인해 일부 시스템은 최신 AI 애플리케이션을 원활하게 구동하지 못할 수 있기 때문”이라고 덧붙였다.

대기업이 AI PC나 코파일럿 PC에 프리미엄을 지불할지, 아니면 해당 방향을 선택할 경우 공급사에 더 저렴한 SKU 제공을 요구할지에 대한 전망도 제시됐다. 안토니아디스는 “AI PC 분야에서 시장 흐름에 맞설 만한 영향력을 가진 기업은 없다고 본다. 기존에 장기 벤더 계약을 체결해 둔 기업이 상대적으로 유리한 위치에 설 것”이라고 설명했다.
dl-ciokorea@foundryco.com

인텔에 따르면 코어 울트라 시리즈3는 글로벌 주요 파트너사의 200개 이상 제품 설계에 탑재되며, 인텔이 지금까지 선보인 AI PC 플랫폼 가운데 가장 폭넓게 글로벌 시장에 공급되는 플랫폼이 될 전망이다.

인텔 클라이언트 컴퓨팅 그룹 총괄 부사장 짐 존슨은 “이번 코어 울트라 시리즈3는 전력 효율과 CPU 성능을 한층 강화하고, 동급 최고 수준의 GPU와 개선된 AI 연산 성능을 제공한다. x86 아키텍처 기반의 애플리케이션 호환성 역시 강화했다”라고 말했다.

노트북에 탑재되는 코어 울트라 시리즈3 모바일 라인업에는 인텔 아크(Arc) 그래픽을 탑재한 새로운 등급의 인텔 코어 울트라 X9 및 X7 프로세서가 포함된다. 이들 프로세서는 이동 중에도 게이밍, 콘텐츠 제작, 생산성 등 고급 워크로드를 동시에 처리하는 멀티태스킹을 위해 설계됐다. 최상위 SKU는 최대 16개 CPU 코어, 12개 Xe 코어, 50 NPU TOPS를 제공하며, 멀티스레드 성능은 최대 60% 향상, 게이밍 성능은 최대 77% 이상 개선됐다. 또한 최대 27시간 지속되는 배터리 수명을 목표로 설계됐다.

코어 울트라 시리즈3 제품군에는 메인스트림급 노트북 구동을 위해 설계된 인텔 코어 프로세서도 포함됐다. 울트라 시리즈3와 동일한 기본 아키텍처를 활용하는 인텔 코어 라인업은 더 저렴한 가격대에서 가성비와 효율성을 갖춘 노트북을 설계하도록 지원한다.

시리즈3 엣지 프로세서는 PC 버전과 더불어 최초로 임베디드 및 산업용 인증을 획득했다. 이는 확장된 작동 온도 범위, 성능, 그리고 24시간 상시 가동이 가능한 신뢰성 등 산업 현장의 까다로운 요건을 충족한다.

인텔은 핵심 엣지 AI 워크로드에서 코어 울트라 시리즈3가 대규모 언어 모델(LLM) 성능 최대 1.9배 향상, 엔드투엔드 비디오 분석에서 와트·달러당 성능 최대 2.3배 개선, 비전-언어-액션(VLA) 모델 처리량 최대 4.5배 향상됐다고 밝혔다. 또한 통합형 AI 가속을 통해 기존의 멀티칩 CPU·GPU 아키텍처 대비 단일 시스템온칩(SoC) 솔루션으로 더 나은 총소유비용(TCO)을 제공한다고 설명했다.

코어 울트라 시리즈3 프로세서를 탑재한 소비자용 노트북은 2026년 1월 6일부터 사전 예약할 수 있다. 해당 제품은 2026년 1월 27일부터 전 세계 시장에서 판매될 예정이며, 추가 제품들은 2026년 상반기 동안 순차적으로 공개될 예정이다. 코어 울트라 시리즈3 기반의 엣지 시스템은 2026년 2분기부터 출시될 예정이다.
dl-ciokorea@foundryco.com