Share this article:

This week’s earnings reports all say the same thing. The money spent on AI work is coming back. Microsoft’s Azure reached $33.17 billion in quarterly revenue, up 48% year over year, and free cash flow of $18.6 billion was more than double the market’s expectations (Global Economic). At Amazon’s earnings call, CEO Andy Jassy brought up the plan for AWS to reach $1 trillion in annual revenue, and the external sale of its own AI chip Trainium (Digital Daily). The day Netflix announced Q3 revenue of $12.874 billion, up 11.8%, and net income of $3.664 billion, up 43.9%, it also announced a cut of 1,000 employees, about 6% of total headcount (News1). The largest cut since 2022, focused on product, technology, legal, and human resources. The cut was not made because results worsened. It was made in the middle of record profits. So I read this week’s news as “the day work was bought back.”

Earnings up, footprint down, AI buying back the work: an image visualizing the concept A visual of the post’s core concept.

The Day Layoffs No Longer Follow Results

Disney’s moves are clearer. It cut about 1,000 people in April, and several hundred each in July and September, three layoffs this year. Over the same period, the streaming division’s operating profit more than doubled year over year to $712 million, and Peacock kept a second consecutive profitable quarter even while shrinking its streaming technology organization (News1). Netflix itself said the layoffs target “overinvested support functions.” The formula “when results worsen, layoffs follow” is no longer a safe formula. These cuts read as a structural choice to make AI and automation investment efficient, not a response to slowing results.

Roland Berger’s analysis supplies the background. Media company hiring data from 2022 to 2025 shows demand for jobs centered on technical execution falling and demand for creative and strategic roles rising. As repetitive work is automated by AI, specialized work is being consolidated. What is shrinking is not cost but the “operating parts,” and what remains is the “judgment parts.” For domestic broadcasters, content producers, and IP companies, this overseas media example becomes a benchmark of “keep the headcount, improve the margin with automation.”

Where Is the Work Going

Disney described its target organization in a job posting itself. A legal and global business organization that automates contract review, legal research, and regulatory analysis with LLMs and AI agents. It is not that lawyers’ work disappears. It is a structure where the repetitive parts of it move to agents.

This scene is not a media industry story alone. With Microsoft’s $18.6 billion free cash flow and the $700 billion this year in data center and AI equipment investment forecast for big tech as a whole (Reuters), the flow of money is converging on one place. Microsoft’s capital spending target for this year is $190 billion, and Q3 spending came to only $45.2 billion due to equipment delivery delays, as investment continues. Starting last month, it began publishing Azure revenue in dollar amounts instead of growth rates, a change that invites direct comparison with AWS and Google Cloud. It is also the point where observers note that hyperscalers carry their own dependency, with OpenAI-related revenue reaching $10 billion. Amazon’s Q3 cash-basis capital spending was $75.9 billion, about 101 trillion won. CEO Jassy went one step further. He sees an era in which 6 to 8 models handle the majority of work within 2 to 3 years, and companies run multiple models together inside one application, on a multi-model operating platform. The point is that the platform has to provide model choice, performance per cost, security, and model swap. The era of a single strongest model is over, and the era of splitting work across models has begun.

The demand side is swelling at the same speed. According to Yonhap News reporting, 34.5% of SpaceX’s Q3 revenue belongs to the anonymous customer “Customer B,” which is Anthropic. The scale now exceeds its previous largest customer, the US federal government. SpaceX’s AI division quarterly revenue doubled from $4 billion to $7.8 billion, and the company raised $40 billion, about 53 trillion won, for AI chip purchases. It is a computing contract of up to $120.5 billion, about 160 trillion won, through 2029, with $3.1 billion flowing per month at full operation. What is interesting is that demand does not concentrate on one. Anthropic is reportedly also holding multiple contracts with other neoclouds such as Enscale and Lambda, keeping open the possibility of exit after 2027. Musk warned that the majority of the more than 200 neoclouds will undergo small-scale decline and consolidation. The computing economy is growing fast, and it is not yet decided who will survive and who will fall out.

The Platform Side Is Already Preparing

Once agents actually start turning work, the center of competition moves from “how well do they do it” to “how do you control them.” That is why SAP announced “Joule Work,” integrating agent development and operations, at TechEd 2026. It supports the choice of more than 70 AI models, connects work inside and outside its own systems through open protocols such as A2A and MCP, and provides an agent gateway where third-party agents can also go back and forth. In particular, the newly introduced enterprise AI gateway manages authentication, policy, routing, and auditing of AI traffic in one place, and even includes a governance assistant for responding to the EU AI Act. It also runs a promotion to deploy agents for free on the managed Joule Studio runtime until December 2026 (VentureSquare). Even SAP, which built the operating system of enterprise work with ERP, lists control as the buying criterion of the agent era.

Security leaders’ numbers back up this flow. In a16z’s survey of 100 CISOs at companies with more than 1,000 employees, 87% forecast increased security spending in 2027, and all seven top priorities were AI-related (Digital Today). By usage, 92% already use generative AI tools, 83% run coding assistants and agents, and 55% connect agents to internal applications through MCP servers. Yet only 18% actually operate AI SOC tools, and offensive security adoption stays at 13%. Purchase plans within 12 months: agent security 45%, offensive security 42%, AI governance 38%. The gap between the will to prepare and operating results is wide. The consulting firm Bain analyzes that AI can compress an attack process that normally takes four weeks to 18 hours. The security industry itself is being reorganized by AI. a16z led the Series B investment (about $255 million) in Armaiden, an AI offensive security startup. The background is the view that existing strong players have not yet caught up in the latest AI areas, and new security companies will rise. Security staffing forecasts split in half between additions and freezes, and the direction is for AI to extend coverage, not replace people. The 1st and 2nd worries CISOs listed are AI exfiltration of sensitive data (47%) and agent privilege overreach and hijacking (46%).

The Same Wave, the Domestic Coastline

The same wave is coming in a slightly different shape domestically. Domestic OTT and broadcasting have different employment conditions, so large-scale layoffs are unlikely to appear immediately, but the efficiency pressure on repetitive work that AI can replace points in the same direction. Discussions are underway on a 1.5 trillion won K-content development budget, and KT and NC AI are expanding into media-specialized AI business (News1). That is the background for growing demand in content production and post-production, legal, and operations automation. The theme of “moving work to systems” is extending beyond IT. The government will start development of a Korean-style sovereign medical AI next year, and plans to connect 72 nationwide responsible medical institutions to a public medical AI highway by 2029 (News1).

The infrastructure side shows overlapping moves. Samsung SDS had Q3 revenue of 3.7725 trillion won and operating profit of 241 billion won, with external cloud driving growth, and external reports say external cloud revenue rose 92% (The Bell). It unveiled a roadmap to expand AI infrastructure to 800MW or more by 2031, and plans to start a B300 GPU service in Dongtan by year-end, then roll out Asia’s first next-generation GPU-based service in March next year. Naver Cloud signed an AI factory deal of up to $9 billion with Brookfield, and will begin operation from 55MW in the first half of 2027. On the longest contract basis, the nominal supply price of 18.74 trillion won is about 12 times Naver Cloud’s revenue last year, making it a business that expects an operating margin of 20% or more at maturity. Telecoms are not stepping back either. KT announced it will invest about 5 trillion won over the next five years to secure 1GW of additional AI data center capacity, but its equipment investment in the first half of this year fell 15.1% year over year, and R&D spending fell 28.2% (BigDataNews). The blueprint has grown, the funding has tightened, so next year’s execution is the test. It is a year in which compute and the work running on it move at the same time.

What Should Companies Prepare

This week’s news leaves pain that sorts into four branches.

Disney’s legal agent scenario is an audit problem. Contract review, legal research, and regulatory analysis touch confidential data and require proof of results. The agent privilege overreach and hijacking that the CISO survey pointed to is a safe execution problem. The era has arrived in which the working agent itself becomes an attack target. As SAP’s gateway lifts authentication, policy, routing, and auditing to standard spec, control has become not an option but a buying criterion. CEO Jassy’s 6-to-8 model forecast and the inference cost that Amazon pointed to as an enterprise adoption barrier are a cost problem. Which model does which work, and at what cost, determines the total cost of agent operations.

ThakiCloud’s answer to these four branches of pain is Paxis. As an Agent-Native Cloud, it has completed the official product v1.1 GA, and its positioning is one. It treats work as a first-class resource. Inside this platform, agents are born on top of Skills, Tools, Policies, and Audit Logs, and do their work there. Autonomy holds permissions in stages from L0 to L3, the policy gate stops dangerous actions before execution, and the audit log leaves everything behind. It runs in isolated sandboxes, extends through MCP connectors and the skill market, and works in sovereign and on-premise K8s environments. CostRouter picks the suitable model for each work. In an era when 6 to 8 models divide up the work, which model and at what cost the work runs on is itself the competitiveness of operations.

Earnings grew and the footprint shrank. What goes into the shrunken space is not people but agents. The question going forward is one. Where will the working agent do its work, under whose permissions, and with what evidence? The company that answers this question first can keep the judgment parts that remain.

References

This post was written by synthesizing the following news.

Share this article:

Tags: agentops, ai-agents, ai-governance, enterprise-ai, multi-model, workforce-automation

Categories:

Updated:

You may also enjoy