20VC: Atlassian CEO on Why Agents Won't Replace Interfaces | Why Enterprises Won't Trust Chinese Models | Why The Jury Is Still Out On The Benefits of Specialised Models with Mike Cannon-Brookes

Mike Cannon-Brookes · Atlassian

Labelling content as AI-touched will fade; quality should not be judged by AI involvement, provenance matters more.

Mike argues that the question of whether content was touched by AI is not simple, and that it is not inherently good or bad. He says watermarking AI text implies AI text is bad, which is not true. The meaningful question is how much a human contributed, such as adding one or two ideas versus generating a 3,000-word essay from a short prompt. He expects people to get used to content that may be entirely synthetic, and to care about the quality of the output, similar to how viewers do not care whether a film has CGI or colour grading. He sees provenance and accountability as important, but as a design problem rather than a label on AI content.

“We see this with the anthropic watermarking features. It implies that watermark text in AI is bad. That's not true. The question is, did I contribute one or two ideas?”at 12:23

Mike Cannon-Brookes · Atlassian

Technology production now outpaces consumption; features ship faster than customers can use them in practice.

Mike argues that the production of technology is now faster than consumption. The industry has reached a tipping point where it can build features faster than customers can absorb them in a way that changes their business. He compares this to a phone that has supercomputer-level power but rarely runs its CPU at full capacity, suggesting there is a limit to how much AI usage will be consumed.

“Right now, the production of technology is faster than the consumption of technology. It is ahead. We've reached a tipping point. We can build features faster than customers can consume them in a realistic way that changes their business.”at 31:08

Mike Cannon-Brookes · Atlassian

Jury is out on specialised models; generalist models are improving faster and eroding most specialised advantages.

Mike says the benefits of building specialised models are unproven. Specialised models do well in some cases, such as those with proprietary data or narrow applications like robotics and world models. He argues most companies building specialised models are not in those cases. Their price-performance edge may hold at shipping but will not hold over time as generalist models improve. He cites startups from two or three years ago that tried to build their own models, whose work has been overtaken by the quality of generic models. He says many companies that claim to have trained their own model are really fine-tuning a Chinese or open-weight base, which is not training from scratch. Atlassian fine-tunes open-weight models for specific purposes but does not claim them as its own models.

“The jury is really out on the benefits of building specialized models at the moment. There are definitely cases where specialized models are doing really well. Most of the people building specialized models, I would argue, are not in those cases, though.”at 21:05

Mike Cannon-Brookes · Atlassian

Broad AI platforms need generic datasets, whereas real applications depend on specialised data.

Cannon-Brookes contrasts the data needs of general-purpose AI platforms with those of specialised uses. Broad intelligence platforms, he says, look for large-scale generic datasets to sell to many providers. He illustrates specialised needs with a self-driving car, which needs a very specific dataset about roads, signs, traffic, cyclists and their dynamics, and does not need general knowledge such as when Albert Einstein was born. The argument is that value comes from domain-specific data rather than generic breadth.

“Those type of companies are looking for large scale generated generic datasets to sell at scale to many other providers... a self driving car needs to understand all of the bits of a road... It doesn't have to know when Albert Einstein was born.”at 39:25

Mike Cannon-Brookes · Atlassian

Enterprises are increasingly scrutinising AI token spend and price-performance, making cost discipline a key tech-company strength.

Mike says large enterprises are increasingly sensitive to cost and care about price-performance, and that their AI budgets are growing. Atlassian now has products that show token spend across the business and in its engineering organisation, so customers can see which models they use, what value they get and where to cut costs. He argues that buying from a vendor that bakes in AI costs is no different from buying a generic or foundation model directly. He says cost discipline will keep arriving, and that managing technology cost, through intelligent model routing and blending per task, is a core strength of great technology companies. Atlassian's FinOps capability lets it cut prices on heavily used features. He also notes that generic model routers are hard to automate and need enough user intent to work well.

“so large enterprises around the world, they're incredibly sensitive to this type of stuff nowadays. They really, really care about the price performance.”at 25:46

Mike Cannon-Brookes · Atlassian

Price-performance of LLM inference will fall sharply because of heavy capital and talent investment, while usage rises.

Mike says model inference is what he thinks will cost a lot today but not in five years. He argues the amount of energy, capital, time and talent going into price per watt of inference is huge. He expects gains to come over years from new chips, better data center design and distillation of foundation models. He does not believe that the progress in LLM price-performance will not bring unit costs down sharply, given the volume of capital and talent. He accepts the Jevons-style point that cheaper inference will drive much more usage, but still expects cost per unit to fall. Where usage is near free, consumption will explode, but as long as there is a waste component and cost, there will be an economic curve that limits usage.

“The overall progress of that, I can't believe doesn't bring our price per performance unit of an LLM, we don't have a good term for this, down at a precipitous rate just because the volume of capital being applied, the volume of brain and talent being applied is so large.”at 30:33

Mike Cannon-Brookes · Atlassian

Agents will change many jobs, but the change is not necessarily negative and frees people for creative work.

Mike says that in enterprise collaboration and business workflows, agents are already helping move work forward in many cases. This lets people focus on the higher-value, creative parts of their jobs. He expects agents to change many jobs, but he does not think the change is necessarily negative, and he says it is good for Atlassian as well.

“agents are going to help and are already helping, I would argue, in a lot of cases, move the work forward to let people honestly focus on higher value creative parts of the job. I think that will change a lot of jobs, but I don't think it's necessarily negative”at 4:36

Mike Cannon-Brookes · Atlassian

Open-weight models face no customer pushback, unlike Chinese-origin models, so Atlassian avoids Chinese models in production.

Mike says some customers push back on so-called Chinese models in some places, but he sees no pushback on open models. Atlassian uses a lot of Nemotron from its partnership with NVIDIA and some Mistral and other open models, and it does not use any Chinese models in production. He notes that many US models are increasingly distilled from Chinese models without saying so, and that enterprises already use Linux, which has around 40% Chinese contributions. He sees the geopolitical question of strategic control as separate from whether a model is technically usable. He argues that enterprises refusing cheaper open-weight models that give better results will face a hard long-term choice, and that companies will fine-tune open-weight models and present them as their own.

“There is certainly some customer pushback in some places to, quote unquote, Chinese models. There is not customer pushback to open models.”at 17:27

Mike Cannon-Brookes · Atlassian

Easy $15m pre-seed rounds reflect venture capital's own choices and are simply the market price.

Harry Stebbings argues that it is now too easy for founders to raise around $15m as a pre-seed round, and then exit early. Cannon-Brookes responds that this is the venture capitalists' doing in a capital-rich society, and that the rounds reflect the market price for these types of things. The argument is brief and gives no further data.

“that's venture capitalist fault. It's a capital society... that's the market price for those types of things.”at 48:55