What AI Companies Want from Cloud Engineers: Analyzing 1,368 Job Postings

If you want to know where AI companies are heading, job listings are far more honest than press releases. Press releases only say what companies want to highlight. Job listings, on the other hand, show the exact positions where work is stalled due to a lack of headcount. They are documents backed by real money.

However, if you read this simply as “AI hiring news,” you miss the real insight. The true value lies elsewhere.

The problems these companies face today will become the problems of ordinary companies in a few years. Kubernetes was no different: once used only by Google, it is now everywhere. So I reclassified these listings based on “what capabilities they require.”

Here is how the data was gathered: OpenAI and Anthropic make their job data publicly accessible via open endpoints, so I pulled the full text of their postings. Google lacks such an endpoint, so I scraped its job search page. Data is current as of September 8, 2026.

① Kubernetes: The Moment Default Settings Break

First, a quick definition. Kubernetes is a tool that clusters multiple servers into a single resource and automatically determines which server should run a given application. It is the de facto standard in today’s cloud industry.

A sentence from Anthropic’s Kubernetes Platform job posting reads:

“We operate at a scale where defaults no longer work.”

What does this mean? Normally, Kubernetes works fine out of the box. But when you scale beyond thousands of servers, default behaviors become bottlenecks. Consequently, these companies modify and rewrite the internal engine itself.

The postings specify exactly what they modify.

Modifying the scheduler directly. The scheduler decides which server runs a given process. The default scheduler places workloads one by one in sequence. However, AI training requires thousands of processes to launch simultaneously. If even one process fails to secure a slot, the rest wait while holding onto allocated resources. If multiple jobs do this at once, a deadlock occurs where no job can proceed.

This creates a need for gang scheduling—an all-or-nothing approach where all tasks are scheduled together or none at all. On top of this comes topology awareness: placing tasks that communicate frequently on physically adjacent servers. If they are spread far apart, network latency consumes most of the compute time.

Tuning the control plane. Kubernetes has a control plane that orchestrates the entire cluster. The apiserver mentioned in the posting serves as the entry point for all commands, while etcd acts as the state store ledger. As server counts grow tenfold or a hundredfold, this entry point and ledger clog first. Job postings phrase this as “finding the next bottleneck before it finds us.”

There is a key takeaway here. Most companies use managed services like AWS EKS or Google GKE, where the cloud vendor manages the control plane for you. The work described in these postings occurs at the point where managed offerings hit their limit. While most organizations may never reach this scale, it remains a fundamental wall everyone hits when scaling far enough.

② Capacity Planning Outgrows Excel

Capacity planning is forecasting server requirements for upcoming periods. In most organizations, this has traditionally been handled in Excel spreadsheets. Yet both companies are hiring for this as a standalone engineering role.

From Anthropic’s Capacity Engineering posting:

“Anthropic operates some of the largest and fastest-growing infrastructure in the industry—spanning multiple accelerator families, multiple CPU families, and multiple cloud providers.”

Accelerators are specialized chips for AI computation, with GPUs being the primary example. The challenge is that there are multiple types, and companies combine resources across multiple cloud providers. Simply determining “what hardware we currently own and where” becomes a major engineering task in itself.

The Infrastructure Capacity Planner role focuses on building predictive models. The targets for forecasting include:

• Accelerators — By chip family
• CPUs — By specification
• Storage — By tier and access pattern
• Egress — By network route

Here, egress refers to data transfer fees incurred when data leaves a cloud platform. Ingress is free, but egress incurs charges—making it one of the most common surprises on a cloud bill.

The term neocloud appearing in job listings is also noteworthy. It refers to specialized emerging providers focused specifically on GPU rental, rather than traditional hyperscalers like AWS, Azure, or Google Cloud. Its appearance in job descriptions signifies a broader landscape of cloud options.

OpenAI’s job posting states the objective explicitly:

“Replace Excel-based workflows with scalable software systems.“

If you currently manage cloud costs in Excel, take this as a signal that the approach is fast approaching its limits.

③ SRE Is Splitting into AIRE

SRE (Site Reliability Engineering) focuses on maintaining service availability and uptime. Pioneered by Google, it is now an industry standard practice.

At Anthropic, however, there is a dedicated team named AIRE (AI Reliability Engineering), with open positions in London, Dublin, and the US.

“AIRE addresses reliability across the entire path from the SDK through the network, API layer, serving infrastructure, and accelerators.”

The top item under core responsibilities is: “Develop SLOs tailored for LLM serving.”

An SLO (Service Level Objective) is a target metric defining acceptable performance—for example, “99% of responses within 0.5 seconds.” Teams frequently use p99 latency, which measures the 99th percentile slowest request out of 100 to ensure tail latencies are not ignored.

However, traditional metrics do not fit AI services like chatbots well, for three main reasons:

• Latency exists in two phases. Time To First Token (TTFT) and Inter-Token Latency (ITL) operate independently. The first character might stream out instantly while subsequent text stutters.
• Workload varies drastically per request. A simple “Hello” versus “Summarize this paper” differs in compute demands by orders of magnitude, making a single metric benchmark unhelpful.
• Batching introduces trade-offs. Grouping requests increases total throughput, but degrades individual response latency.

Consequently, defining performance using a single target metric is difficult, which likely explains why a distinct specialization emerged.

④ Hardware Management Has Become Software Engineering

A line from OpenAI’s Fleet Hardware Health posting illustrates this shift. A fleet refers to the company’s entire collection of operational servers.

“A single hardware issue can cause significant downtime. As supercomputers grow larger, these risks scale exponentially.”

Single-server failure is routine in standard cloud architecture—you simply isolate or drain the node. In AI training, however, thousands of nodes operate as a tightly coupled system. If one machine falters, the entire cluster halts. Proactively detecting hardware anomalies has thus become a primary task for software engineers.

The Datacenter Server Lifecycle role at Anthropic manages every stage of a server—from receiving and racking to operation and repair—through software. Notably, security is an integral component rather than an afterthought.

“Ensuring that every machine in the fleet operates on a trusted, attested, hardware-verified chain of integrity is central to this role, not an ancillary detail.”

Attestation is the automated verification that a physical server is genuine and untampered with. While many organizations treat hardware verification as a periodic compliance activity, here it functions as a core operational baseline.

Notably, this domain aligns with shifts in troubleshooting methodologies. A wave of recent research has focused on silent data corruption in large GPU clusters—where hardware returns incorrect compute results silently without throwing errors.

The Three Companies Are at Different Stages

These patterns reflect the software layer. Looking at physical infrastructure and facilities, the three companies diverge significantly.

OpenAI is in active build mode. A substantial portion of the 114 listings in its Scaling org center on facility construction.

• Data Center Infrastructure Architect ($360K–$530K) — Designs power, cooling, and equipment layouts, simulating them via digital twins (virtual replicas used to model real-world thermal and operational scenarios).
• Data Center Controls Network Engineer ($257K–$327K) — Focuses not on data server networks, but on operational technology (OT) networks connecting chillers, switchgear, and power distribution units.
• Datacenter Networking Technician ($157K–$221K) — On-site 5 days a week in Abilene, Texas, running optical fiber cabling and testing connectivity.

A key trend emerges here: server IT networks and facility OT networks are converging. Historically managed by separate teams with separate equipment, AI data centers cannot maintain this boundary because cooling failures instantly shut down compute clusters.

Anthropic is expanding globally. Many of the 25 roles in its Compute org are region-specific country leads.

Compute Country Lead, Korea      (서울)
Compute Country Lead, Japan      (도쿄)
Compute Country Lead, Canada
Data Center Energy Lead, EMEA    (런던)
Data Center Energy Lead, Australia (시드니)

There is a role based in Korea. The posting outlines responsibilities spanning site selection, construction, and operation, stating the candidate is “ultimately responsible for bringing gigawatt-scale compute online in Korea.” Another posting mentions that “Anthropic is investing 50 billion USD ($50B) in US computing infrastructure.”

Google is in steady-state operations. Searching for "data center" yields 20 results on the first page, all focused on field maintenance. The requirements for Data Center Technician II illustrate this emphasis:

• Linux experience and 2 years of server hardware troubleshooting
• Ability to lift 50 lbs (23 kg), bend, kneel, and climb ladders regularly

This stands in contrast to OpenAI’s $360K architect roles within the same sector. Having operated data centers for over two decades, Google’s current hiring focus is on operators rather than initial system architects.

All Three Companies Are Designing Custom Silicon

One unexpected finding emerged: while Google is known for custom silicon, counts show that all three companies are actively hiring semiconductor engineers.

Chip development follows distinct phases: RTL design specifies logic in code; design verification (DV) tests logic correctness; and bring-up powers on and validates initial silicon samples.

• OpenAI (12 postings) — Openings cover design, DV, packaging, and silicon bring-up, alongside specialized semiconductor legal counsel.
• Anthropic (4 postings) — Roles include Silicon Engineer and research positions exploring reinforcement learning for chip design.
• Google — With multiple generations of custom TPUs in mass production, hiring focuses heavily on iterative hardware engineering.

How does this impact cloud engineering? Silicon heterogeneity leads to fragmented hardware fleets. This explains why capacity planning roles emphasize “multiple accelerator families” and why custom Kubernetes scheduling becomes necessary, as infrastructure can no longer assume identical GPU nodes.

Common Factors Across All Three: Power and Community

Power. All three hire energy leads. Anthropic’s postings describe securing hundreds of megawatts directly from grid operators and regulatory bodies.

A shared theme across all three companies is that power availability—rather than GPU supply—is the primary bottleneck.

This has broader implications: future GPU availability in specific regions may be constrained by local power grid capacity rather than hardware supply chains.

Community. Anthropic is hiring a Community Engagement Manager, Data Centers across Texas, Australia, and Canada. The posting describes this as a “relationship role, not PR or government affairs,” tasked with identifying local resident concerns proactively. Google maintains similar roles, reflecting how data center siting has become a focal point for community relations.

How Current Learning Maps to Emerging Roles

In closing, one key point stands out:

These four shifts are not limited to AI companies. As server counts expand, control plane performance becomes a bottleneck for any organization. As spend grows, capacity planning must move beyond spreadsheets.

These job postings represent a list of problems encountered first by organizations operating at extreme scale. Even if you are not planning a career move, studying them offers a window into the infrastructure challenges mainstream engineering teams will face in the coming years.

A direct indicator for regional markets is the opening for Compute Country Lead, Korea. Roles like this typically precede subsequent hiring in local power infrastructure and facility operations.

⚠️ The job posting data referenced here was pulled on September 8, 2026 from public endpoints (OpenAI via Ashby API, Anthropic via Greenhouse API, and Google via job search results). Quotes reflect original English job descriptions. Salary figures represent US base pay as listed, excluding equity. Job listings change frequently. Google does not disclose total listing counts, so figures represent page 1 search results. Interpretations regarding AI latency characteristics and managed Kubernetes limits reflect the author’s analysis rather than official company statements.

Frequently Asked Questions

I am just starting to learn cloud engineering. Do I need to know all of this?

Not immediately. The challenges described here arise at a scale of thousands of servers, which most organizations do not reach. However, understanding these trends helps guide what to study deeper. For instance, going beyond basic Kubernetes usage to understand scheduler internals, or understanding why egress charges exist on cloud bills, are valuable concepts you can integrate into your current learning.

If I use AWS EKS, does that mean I won’t ever need to modify the scheduler?

In most cases, yes. Managed services exist specifically to abstract away control plane complexity. However, it is valuable to know where managed services hit their limits. Signals include workloads waiting indefinitely for allocation, partial job deployments stalling, or control plane command latencies spiking. Recognizing these symptoms as scale boundaries rather than generic application bugs changes how you troubleshoot.

Could you explain gang scheduling again in simpler terms?

Consider a group of 10 people arriving at a restaurant. If only 6 seats are available and 6 people sit down, the remaining 4 wait standing up, while the seated 6 cannot start their meal. If another group does the same, both groups lock up partial seating and nobody eats. Distributed AI training behaves identically: all thousands of tasks must acquire slots before execution begins. Gang scheduling enforces an “all tasks get scheduled together, or none do” rule.

Is capacity engineering the same as cloud cost management (FinOps)?

They overlap, but their focus differs. FinOps emphasizes analyzing and optimizing historical spend. Capacity engineering focuses on forecasting future demand and allocating resources programmatically through custom software rather than spreadsheets—as captured by OpenAI’s posting specifying the replacement of Excel workflows with software systems.

Do I need an AI background to work on AI company infrastructure?

You do not need to build AI models yourself. Most infrastructure roles do not directly touch model architectures. However, you must understand how AI workloads differ from conventional software—such as why thousands of tasks must spin up concurrently and why node proximity matters. This knowledge builds directly on cloud engineering principles rather than machine learning theory.

Will similar roles open up in regional markets?

On the facilities side, signs are already emerging with listings like Compute Country Lead, Korea in Seoul. Software infrastructure roles usually follow with a slight lag. Because the four shifts described here are functions of scale rather than AI specifically, high-scale regional platforms are likely to introduce similar engineering specializations over time.