The hard parts are the new moat

Read a month of venture news in one sitting and the pattern is hard to miss. The easy parts of building a company kept getting cheaper. The hard parts kept getting more valuable.

Model prices fell again. Coding tools now write more code than most teams can review. New apps launch every day. Meanwhile chips stayed scarce, power came up in every data center deal, large companies got stricter about which models they allow, and the teams that could prove their results pulled ahead.

Founders who build around the hard parts will have the easier time defending what they build.

Cheap software cuts both ways

Anthropic released Claude Haiku 5.5 on October 7, priced 90% below its predecessor for prompts of up to 100,000 tokens (Anthropic). That fits a long trend. Sam Altman wrote in early 2025 that the cost of a given level of AI falls about tenfold every year (Three Observations), and the Stanford AI Index measured a more than 280-fold drop in the cost of GPT-3.5-level performance between late 2022 and late 2024 (AI Index 2025). Products that made no economic sense a year ago now do.

Your competitors get the same discount. Developers shipped around 600,000 new apps to Apple's App Store in 2025, 30% more than the year before (9to5Mac, citing Sensor Tower). People's free time did not grow by 30%. If the only thing your product has going for it is that it works, expect a copy before your next round.

What holds up is anything that improves with use. Data nobody else has, feedback loops that sharpen the product with each customer, and workflows built into how a customer actually operates all qualify. Design and a clear point of view count for more than they used to, because they are still slow to copy.

Value is moving toward hardware

Cheaper tokens have increased demand for compute. Nvidia reported $96.2 billion in revenue for the quarter that ended in July, up 106% on the year, with $89 billion of it from data centers (Nvidia). Jensen Huang's summary was "compute is revenue." The limits are now physical. The International Energy Agency expects data center electricity use to more than double by 2030, to around 945 terawatt-hours (IEA). Huang also calls AI and robotics Nvidia's two largest growth opportunities (CNBC).

The credit market's nervousness about GPU cloud providers this month is worth reading closely. In mid-October a CoreWeave bond briefly yielded more than 15% and its shares fell over 15% in a day (SiliconANGLE). Gartner now expects 80% of independent GPU clouds to fail within four years (Bisnow). Nobody doubted the demand. The worry was debt, and whether expensive chips pay back before newer ones replace them, a risk Sequoia's David Cahn flagged back in 2024 (AI's $600B Question). Hardware founders need engineering depth and financial discipline in equal measure, and payback period belongs on the first page of the plan.

This is the core of the Blue Trail thesis. Photonics, semiconductors, robotics and energy are difficult to build, slow to copy and increasingly central to how AI gets delivered.

Proof is the bottleneck

AI improves fastest where results can be checked automatically. Andrej Karpathy put it plainly last November: software "easily automates what you can verify" (Verifiability). In math and code, a model can try an answer, test it and try again without anyone leaving the computer. Google DeepMind's AlphaEvolve scores its own programs this way, and one scheduling heuristic it found now recovers 0.7% of Google's worldwide compute (Google DeepMind). Biology and manufacturing still wait on the lab. As Dario Amodei notes, cells and animals "run at a fixed speed" (Machines of Loving Grace).

For deep tech teams, the speed of that loop is a competitive advantage. Simulation, digital twins, automated test rigs and well-instrumented pilots can turn a quarterly learning cycle into a weekly one.

The same applies to AI agents, where small errors compound over many steps. In May, METR found that a leading model completes tasks of about twelve hours half the time, but only tasks of about ninety minutes when it has to succeed four times out of five (METR). In an earlier METR review, none of the agent-written code changes that passed automated tests could be merged as they were (METR). Most of the value sits around the model, in the data, the tools, the memory and the evaluations. Domain experts who can define what a correct result looks like will end up with more leverage than whoever writes the prompts.

Selling to software, and to cautious enterprises

On some developer platforms, agents already do much of the work. Coding agents trigger more than half of deployments on Vercel, up from under 3% six months earlier (IT Brief), and agents create over 80% of new databases on Neon (Databricks). When an agent chooses the tools, your documentation and your API do the selling.

Large companies are heading in a different direction. Enterprise spending on generative AI reached $37 billion in 2025, more than triple the year before, and three quarters of use cases are now bought rather than built (Menlo Ventures). Buyers run more models than they used to, with 37% using five or more, and security and cost weigh more heavily in which ones they pick (a16z). They route each task to the cheapest model that can do it, and they keep models they cannot vouch for out of production.

So go-to-market now has two audiences. Agents need clear documentation, predictable pricing and an API they can use without a human in the loop. Enterprises need strong security, control over their data and a straight answer about which models sit behind your product. Point solutions that serve neither will struggle to stay independent.

Choose your funding path early

Money is concentrating. Funds of $500 million or more took 78% of the US venture capital raised this year while making up 6% of funds, and only 211 firms have closed a fund, against 927 in 2022 (PitchBook-NVCA, via SiliconANGLE). AI accounts for a record 82.7% of deal value. At the top end, some seed rounds now pass a billion dollars, as Yann LeCun's AMI Labs showed in March (Tech.eu). Mid-sized funds are being squeezed, and a profitable company growing fast at modest scale can struggle to get a meeting with a fund that needs billion-dollar positions.

Decide early which kind of company you are building. Venture scale comes with large rounds and expectations to match. Capital efficiency needs investors whose check sizes and return targets fit that path. Both can produce excellent businesses. Problems start when founders raise for one and build the other.

Read your documents, too. This month's public dispute over a pro-rata waiver during a Series B is a good reason to know exactly who can waive what in your term sheet (Venture Awaits). Adviser and board-observer arrangements belong in writing.

Six things to do this quarter

1. Name what gets better every time a customer uses your product. If you can't, put it on the roadmap.
2. Set a budget for AI tools. Uber used up its full 2026 AI coding budget in about four months and now caps spending per engineer (TechCrunch).
3. Make your documentation and API easy for agents to use, and check how agents rank you against competitors.
4. Measure how long it takes to learn whether a change worked, then work on cutting it in half.
5. Review pro-rata, adviser and observer terms before your next round, not halfway through it.
6. Treat security as part of the product. Assume a breach will happen, use AI to review code continuously, and start listing what will need post-quantum encryption.

If you are working on AI innovation, Blue Trail would like to hear from you.