Horowitz Andreessen Academy Trades Degrees for Portfolios as AI Thins Entry-Level Jobs

Andreessen Horowitz on Tuesday launched the Horowitz Andreessen Academy, a for-profit, full-time school in San Francisco for high school graduates who would rather build than enroll in a four-year university. The a16z-incubated academy has raised $42 million, $35 million of it from the venture firm, and opened applications for a tuition-free, roughly 50-student Founding Class Fellowship that begins in September 2027.

The school arrives at a specific moment in the AI labor market. Stanford's Digital Economy Lab reported in August that employment among 22- to 25-year-olds in the most AI-exposed occupations now sits about 19% below where it would be had it kept pace with less-exposed peers, a gap driven mainly by companies hiring fewer young workers. The academy's ten founding partners include Anthropic, OpenAI, Google and Meta, the companies whose models are reshaping that entry-level work. They are now helping design a pipeline that replaces the degree with a portfolio and a network, although, by the school's own account, none of them has committed to hiring a set number of graduates.

That makes the academy less a college replacement than an attempt to rebuild the junior-talent on-ramp that AI is wearing away. Early-admission applications are due November 5, 2026.

The Horowitz Andreessen Academy's Offer to the Class of 2027

The academy is led by founder and CEO Gagan Biyani, who co-founded the online-course marketplace Udemy and later Maven, a cohort-based learning startup that raised $25 million in a round led by a16z. Marc Andreessen and a16z general partner Erik Torenberg join Biyani on the board. Beyond a16z, the $42 million came from individual investors including Quora co-founder Adam D'Angelo, Shopify CEO Tobi Lütke, DoorDash CEO Tony Xu, Fidji Simo, recently OpenAI's CEO of Applications, Y Combinator's Garry Tan, Alpha School's Joe Liemandt and Palantir CTO Shyam Sankar.

The one-year fellowship follows a fixed calendar. Students spend September through April building projects on a San Francisco campus, travel abroad in May, and work in a co-op placement inside a technology company from June through August. The work is open-ended, from starting a company to training a model or writing a book, with progress presented to peers and practitioners every week.

Each student receives more than $50,000 in compute credits and technology resources from partners, plus a $5,000 travel and exploration budget. The ten founding partners are Anduril, Anthropic, Coinbase, Google, Meta, NVIDIA, OpenAI, Palantir, Replit and Stripe, and the school says more than 40 additional companies, including Databricks, Cognition, Notion and Applied Intuition, have signed on as hiring partners.

The homepage features instructors and mentors including Sam Altman, Jensen Huang, Satya Nadella, Mira Murati, Fei-Fei Li, Ali Ghodsi, Dylan Field, Travis Kalanick and Naval Ravikant, with roughly 240 more names listed in the broader network. Ben Horowitz framed the rationale in industrial terms, saying the training built for the Industrial Revolution "isn't going to map perfectly onto the AI revolution."

Proof-of-Work Admissions Replace Essays, Not Transcripts

Admission is weighted toward what applicants have "built, shipped, or earned." The first-stage application asks for a personal portfolio website, a short video, contact details, test scores and a high school transcript. There are no essays or recommendation letters at that stage, although transcripts and scores remain required for the record.

Eligibility is defined by education rather than age. Applicants need a high school diploma or equivalent by August 2027 and become ineligible if they will have completed more than one year of full-time college by then. Students under 18 may apply with a parent or guardian co-signing, and U.S. citizenship is not required, though the school says it cannot guarantee visa approval. The launch release separately says most students will arrive directly from high school or "after 1-2 years at an elite university," a looser framing than the formal one-year cutoff that prospective applicants should check directly.

The school showcases early applicants to signal the bar: one designed and built a six-axis robot arm, another built an AI book indexer with six-figure annual recurring revenue, a third built a sports app used by 83,000 players, and a fourth raised $100,000 for an enzyme-engineering startup. These are the school's self-selected, unaudited examples of its target profile.

Inside the AI Curriculum: Inference, Evals and Agent Harnesses

Courses run from two days to four weeks and are taught by practitioners rather than academics. The published directory lists 62 courses and tracks. Beyond startup staples such as founder-led sales and cap tables, the AI-specific offerings track where applied AI engineering has moved over the past two years.

One course, AI Inference Engineering, addresses the serving side of language models: how to run a trained model cheaply and quickly in production. That work spans batching many user requests onto one GPU, caching intermediate attention computations so repeated context is not recomputed, quantizing model weights to lower-precision numbers to cut memory use, and choosing hardware against latency targets. For most startups built on foundation models, inference cost rather than training cost determines gross margin.

Another, Building and Evaluating AI Products, is taught by Hamel Husain and Bryan Bischof, practitioners known for work on LLM evaluation. Evals are the test suites of AI software. Because a model's output varies from run to run, teams build graded datasets, automated judges and error taxonomies to measure whether a change actually improved behavior. Without them, a product demo can look strong while failing silently on the long tail of real inputs.

A third, Harness Engineering: Designing the Operating Layer for Reliable Agents, targets the scaffolding around a model: the tool definitions, memory, context management, permission checks and retry logic that turn a chat model into an agent capable of multi-step work. In 2026 agent reliability depends as much on this layer as on the underlying model, which is why labs and developer-tool companies now ship their own harnesses.

Read more: NVIDIA open-sources an autonomous agent harness optimizer

At the research end, an AI Research & Models track lists Fei-Fei Li and Yann LeCun among its instructors, and a Compute, Infrastructure & Security track includes Daniel Gross, Naveen Rao and Palo Alto Networks CEO Nikesh Arora. The curriculum treats AI as a building material to be deployed, evaluated and scaled rather than a subject to be studied from first principles. That choice reflects the school's premise, stated in its a16z announcement, that teenagers can now "envision a new business in the morning and earn revenue by sundown."

The $50,000 Compute Allowance Equals Months of H100 Time

The $50,000 compute allowance is the most concrete AI resource in the package, and it is worth sizing. On-demand rental of an NVIDIA H100, still the workhorse accelerator for fine-tuning and inference, ranged from roughly $2 to $4 per GPU-hour at specialist clouds and some hyperscalers in 2026 price surveys. At those rates, $50,000 buys about 12,500 to 25,000 H100-hours, or roughly two to four months of one eight-GPU server running continuously.

That is serious capacity for a student. It supports repeated fine-tuning of open-weight models in the tens of billions of parameters, reinforcement-learning experiments on small models, training compact robotics or vision policies, and heavy API usage for agent products. It is nowhere near frontier pretraining, which runs on clusters of tens of thousands of GPUs. The credits also come "from Academy partners" and are bundled with "technology resources," so the realized value likely depends on each partner's terms, platform and expiration rules, which the school has not detailed.

The allowance also functions as distribution. A student who builds a first product on a partner's cloud, model API or developer platform during an intensive year is a likely long-term customer. Startup-credit programs at cloud providers and a16z's own speedrun accelerator, which bundles more than $5 million in credits per company, use the same logic.

AI Labs Co-Design a School as Entry-Level Hiring Shrinks

Much of the launch-day coverage centered on Silicon Valley's dropout mythology. The labor data points to a more practical driver. The Stanford Digital Economy Lab's updated "Canaries in the Coal Mine" paper, using ADP payroll data through June 2026, found no economy-wide AI job displacement but a widening gap for young workers in exposed occupations such as software development and customer service. In levels, employment for 22- to 25-year-olds in the two most exposed quintiles fell about 11% from November 2022 to June 2026, while the same age group in the least exposed quintiles grew about 10%. The researchers attribute the gap primarily to reduced hiring of young workers rather than layoffs, and they find the declines concentrated where AI substitutes for tasks rather than complementing workers.

That finding describes a broken rung. The classic path into tech, a computer-science degree followed by several years of learning on the job at a large company, depended on firms paying juniors to do work that AI tools now handle. Erik Brynjolfsson, the paper's lead author, has argued that employers will need to train young people more explicitly instead of expecting them to absorb skills through routine work.

The academy offers a private version of that explicit training, with the co-op standing in for the first job. Its founding partners are also its prospective employers, and several are the AI developers whose tools are changing what junior work looks like. Biyani told Fortune that partner commitments "vary" and that hiring partners are committed to interviewing or considering students, not to hiring a set number. The school confirms it guarantees neither a co-op placement nor a full-time job.

The labor evidence is not settled. A 2026 analysis by Iscenko and Millet found that junior and senior job postings in highly AI-exposed occupations have declined roughly in parallel since spring 2022, which challenges the claim that AI specifically targets entry-level roles. Either way, the traditional entry path has narrowed, and the academy is positioned to capture candidates who would otherwise compete for fewer junior openings.

Read more: Why verification now defines AI-built software

Between the Thiel Fellowship and Y Combinator, With a Price Tag

TechCrunch described the academy as a cross between Y Combinator and Peter Thiel's fellowship. The comparison is useful mainly for what it shows about money flows. Founded in 2011, the Thiel Fellowship pays 20 to 30 young people a $200,000 grant over two years to skip or leave college, and its alumni include Figma CEO Dylan Field, now listed among the academy's instructors. Y Combinator invests in companies that already exist. a16z speedrun, the firm's own accelerator since 2023, invests $500,000 for 10% upfront plus another $500,000 in a later round and says it has backed more than 250 companies.

The academy sits upstream of all three. It targets people before they have a company, takes no equity, and says students keep 100% of the intellectual property they create. It also takes no income-share agreement and holds no right of first refusal, and a16z has no special claim on student startups beyond the obvious introductions. The planned two-year program reverses the Thiel model: instead of paying students to leave college, the academy expects to charge tuition "similar to that of elite private universities" beginning in fall 2028, pending regulatory approvals.

For a16z, the academy extends a talent funnel that already runs from media and podcasts through speedrun to seed and growth rounds. The notable shift is timing. Venture firms have historically found founders after incorporation; this program meets them at 17 or 18, inside a network the firm organized. The presence of Garry Tan, who leads Y Combinator, among the individual investors suggests the wider founder-sourcing ecosystem views the academy as complementary rather than hostile, at least for now.

Tuition-Free Is Not Cost-Free: The Economics of a For-Profit Academy

The founding year charges no tuition, but students pay their own San Francisco housing, food and living costs. The school cites typical rates of $1,600 to $1,800 per month for a shared dormitory room and $2,300 to $2,600 for a single. Over twelve months, housing alone would total roughly $19,200 to $31,200, before food. Need-based scholarships for living expenses "may be available." The program does not participate in federal or state financial aid, and 529 college savings cannot be used tax-free.

The school's business model rests on the planned two-year program. Biyani told Fortune that higher education is a $2 trillion industry and that there are "lots of ways to turn this into a successful business," adding that the school will see over time how students finance it. He has also argued that for-profit status "forces us to deliver to the customer." Spread across the 50 founding seats, the $42 million raise equals roughly $840,000 per student. That figure is not a spending rate, but it shows how far the first cohort is subsidized relative to any tuition-paying future.

The pricing test is therefore central. The academy has to prove that the most capable teenage builders, and their parents, will pay elite-university tuition for a program with no degree, in exchange for mentors, a peer network, project time and a faster route into technology companies. Biyani has said college remains the right path for most students and that the academy is designed for a small, specific group.

No Degree, No Guaranteed Job: The Limits in the Academy's Own Terms

The school states plainly that it is not accredited, and the fellowship awards no degree, no certificate and no transferable college credit. Students leave with a portfolio intended to demonstrate their work to employers, collaborators and investors. Co-ops are described as "typically" paid in Silicon Valley, but pay is not guaranteed, and hiring partners run their standard interview processes.

Mentor access is also narrower than the headline roster suggests. The school says some mentors will hold office hours only a few times a year and that students must develop relationships themselves after the school makes introductions.

The model carries concentration risk as well. If a partner's hiring slows, if the brand falters, or if AI tools keep compressing the value of early-career technical work, students hold no fallback degree. A student who defers a college admission to attend retains that option only if the college's deferral policy allows it, which the academy says it cannot control.

Admissions Yield, Partner Commitments and 2028 Approvals Are the Next Tests

Three milestones will test the model over the next two years. The first is admissions yield: early applications close November 5, with second-stage decisions by late January 2027, and the number of admits who defer elite-university offers will show whether the value proposition lands. The second is partner specificity. The school has not published what each founding partner contributes, and binding co-op or hiring commitments from Anthropic, OpenAI, NVIDIA or Google would change the offer materially. The third is regulatory: the two-year program depends on approvals targeted for fall 2028, and the path it takes will determine whether the academy can ever award a recognized credential.

The academy's bet is that in an AI-reshaped labor market, demonstrated building ability and proximity to the companies doing the hiring will outweigh a four-year degree for a small tier of students. The Stanford data suggest the old entry path is narrowing, which supports the premise. The unanswered part is who carries the risk. For the founding class, the academy and its investors absorb the tuition. Once the two-year program charges elite-private prices without a degree or job guarantee, that risk shifts to 18-year-olds and their families, and the partners' willingness to convert co-ops into offers will decide whether the trade was worth it.