How to get a job at Anthropic before its IPO

What's actually happening at Anthropic right now?

Anthropic is reportedly struggling to lock down enough cloud compute, because the providers it buys from increasingly see it as a competitor — and this is happening while its infrastructure spending balloons ahead of a planned November IPO. That's the story that broke in late October 2026, and it matters to you as a candidate far more than it matters as gossip. A company that can't get compute as fast as it wants it hires people who can fix that: capacity planners, infrastructure engineers, partnership leads, and the finance and procurement folks who negotiate multi-year commitments. Constraints create headcount. Read the constraint, and you've read the hiring plan.

Pair that with the listing itself. A company heading toward a public offering tightens up everywhere a prospectus gets scrutinised: accounting, internal controls, security, compliance, trust and safety, legal. These aren't glamorous AI jobs, and that's precisely why they're winnable. The queue for a research scientist role is enormous. The queue for someone who has actually run SOX readiness at a fast-scaling company, or built data-centre capacity models, is short. If your background sits in one of those lanes, the pre-IPO window is the single best moment to apply, because the need is urgent and the alternative is an expensive contractor.

Be careful with the specifics, though. IPO timing slips constantly, terms leak half-formed, and a reported plan is not a filed one. Don't write a cover letter that says "congratulations on your November IPO" as if it's done — you'll look like someone who reads headlines rather than companies. Reference the direction of travel instead: rapid infrastructure scaling, growing enterprise demand, the operational maturity a public company needs. That framing stays accurate whether the listing happens in November, slides to spring, or changes shape entirely. Accuracy is a hiring signal in itself at a lab that publishes about honesty.

Which Anthropic roles are expanding fastest?

The roles growing fastest are the ones that touch compute: inference infrastructure, training infrastructure and performance, capacity planning, site reliability, plus the commercial and partnership roles that secure and defend access to GPUs and TPUs. Alongside those sit the safety-adjacent functions — alignment and safeguards engineering, model evaluations, trust and safety, threat intelligence — and the enterprise go-to-market machine selling Claude into large organisations. If you've been reading Anthropic's careers page as a wall of research-scientist listings, look again. The research roster is the smallest, most competitive slice. The infrastructure, product engineering, security and commercial listings are where the volume actually is.

There's a second tier worth your attention because almost nobody targets it: the operational spine of a company about to be public. Technical accounting, revenue operations, procurement, data-centre and facilities programme management, FP&A with cloud-cost modelling, internal audit, privacy counsel, corporate security. None of it requires a machine-learning PhD. All of it requires evidence you've handled scale and scrutiny before. An FP&A analyst who can show they modelled a $200M cloud commitment and caught a pricing error is directly relevant to a company whose largest line item is compute. Say that in your first bullet, not your fourth.

Then there's the applied layer — solutions architects, forward-deployed engineers, developer relations, support engineering for the API and enterprise plans. These roles need people who can read a customer's codebase, debug a prompt pipeline, and explain token costs to a CFO without condescension. If you've spent the last year genuinely building with large language models rather than just talking about them, this is where that experience converts. Ship something public: an evaluation harness, a retrieval pipeline with measured latency, an agent that does one job reliably. A link to working code in your CV header does more for an applied role than any adjective you could write about yourself.

Which keywords from Anthropic's job ads belong on your CV?

Mirror the nouns, not the adjectives. Anthropic's postings lean on concrete technical vocabulary — inference optimisation, training infrastructure, model evaluations, alignment, safeguards, capacity planning, accelerator utilisation, distributed systems, Kubernetes, Terraform, TPU and GPU fleets, observability, RLHF, red-teaming, responsible scaling — and on behavioural language drawn from its own published values: acting for the global good, being honest about uncertainty, doing the simple thing that works. The technical nouns belong in your experience bullets and a skills block. The values language belongs in how you describe decisions you made, never as a quoted slogan pasted into your summary.

Here's the test that separates tailoring from stuffing: every keyword you add must be attached to something you did, with an object and ideally a number. "Capacity planning" alone is noise. "Built the quarterly GPU capacity model for a 4,000-accelerator fleet; cut idle reservation spend 18%" is a keyword earned. If you can't attach a keyword to a verifiable action, leave it off — a skills section stuffed with terms you can't defend in a 45-minute technical screen is worse than a short honest one, because the interviewer will pick exactly the term you were bluffing. Interviewers at a research lab are unusually good at that particular move.

Density matters less than placement. Parsers and human skimmers both weight the top third of page one, your most recent role, and section headings. So put the role-matching language where it lands: a one-line headline naming the job family you're applying for, a tight skills block with real tools, then recent experience bullets carrying the rest. If you want to see which of a specific posting's terms your current CV is genuinely missing, paste the job description into a free CV analysis — the analysis lists the missing skills and generates ATS-friendly rewrites that mirror that job's keywords, which beats guessing by hand at 1am.

How do you survive the first automated screen?

Clear the screen by sending a single-column, text-based PDF with conventional headings, standard date formats and no layout tricks — then confirm the text actually extracts. Most rejections at this stage aren't judgments about you; they're extraction failures. Two-column templates interleave unrelated lines when parsed. Icons replace the words "email" and "phone" with nothing. Skills rendered inside text boxes or graphics vanish entirely. Dates written as "'23–'25" confuse date parsing and can make a two-year role read as a gap. The fastest self-check costs nothing: open your PDF, select all, copy, paste into a plain text editor, and read what comes out. That garbled text is your CV as the system sees it.

Volume discipline matters as much as formatting. A high-signal company like this is receiving an extraordinary number of applications, and the ones that convert are narrow and specific: one posting, one tailored CV, one short note that names the team's actual problem. Applying to eleven different Anthropic roles in a week is a visible negative — recruiters see the full application history, and a scattergun pattern reads as someone who wants a logo rather than a job. Pick two roles maximum, at most adjacent ones, and spend the saved time on the referral and the work sample instead.

Referrals still move the needle more than anything else you can do in an afternoon. Find people doing the work — on the research blog, in open-source commits, at conference talks, in public papers — and send something useful rather than a request. A short, specific message about an eval you ran, a bug you found in a public repo, or a benchmark you reproduced gets replies. "Can you refer me?" from a stranger does not. If you've been out of the market a while and your CV reflects an older era of formatting, it's worth rebuilding it rather than appending to it before you start reaching out.

Is pre-IPO equity worth joining for right now?

Treat pre-IPO equity as a real but illiquid bonus, not as the reason you take the job. Joining shortly before a listing usually means a strike price or grant value set against a very high recent valuation, a standard four-year vest with a one-year cliff, and a post-IPO lock-up — commonly around six months — before you can sell anything. So the money you're negotiating over today is money you might first touch 18 months from now, at a share price nobody can promise you. The honest framing: you're being paid in cash plus a lottery ticket with reasonable odds and a long fuse.

Ask the questions that actually determine the outcome. Are you getting RSUs or options? If options, what's the strike and the exercise window if you leave? What's the current preferred or last-round valuation your grant was priced against, and how many shares are outstanding, so you can compute a percentage rather than admire a headline dollar figure? Is there double-trigger acceleration? What's the lock-up, and are there scheduled trading windows after it? A recruiter who won't answer the share-count question is telling you something. Pre-IPO grants are the most common place candidates mistake a big number for a big outcome.

My contrarian take: the strongest reason to join a lab mid-compute-crunch isn't the equity at all — it's that scarcity teaches you things abundance never will. Engineers who learned to squeeze throughput out of constrained accelerator fleets, negotiated supply when vendors held the leverage, or built evaluations that caught real failures carry skills the whole market will pay for in five years, listing or no listing. Weigh the offer on the work, the team and the compensation floor in cash. Let the equity be upside. If it's the only thing making the offer attractive, the offer isn't attractive.

Frequently asked questions