Issue #10 | August 17, 2026 | Growth stopped meaning what it used to mean.
The math changed and nobody sent a memo. For two years the AI story was about what the models could do. This week, across a dozen shows, the story quietly became about what they cost, and who is actually willing to pay for it. Canva cut its own growth forecast because serving AI features got too expensive. A venture partner told legacy software CEOs to get to AI or accept being worth three times revenue instead of ten. A model that used to be a punchline is now the cheapest way to get frontier performance. None of that is theoretical for a business my size. It is the exact question I ask every time I look at a TPG invoice for API usage.
Now, the week.
Canva told investors its 2026 growth rate is coming down from 30 percent to 20 percent, and the reason was not fewer customers. It was the cost of running AI image generation for all of them. Subsidizing users with expensive frontier model calls is a trap: the more people love the feature, the faster it destroys the unit economics. Canva is reportedly racing to build cheaper in-house models to escape the bill, a process that takes twelve to eighteen months even when it works.
Eric Vishria, a longtime Benchmark partner, put a number on the same problem from the investor side of the table. His message to legacy SaaS CEOs still treating AI as an evening project: get to AI, or be worth three times revenue instead of ten. He was blunter than that in the room, telling founders that every day they hit their old plan on schedule, they are quietly destroying equity value, because the plan itself is now the risk.
For a business my size, the lesson is not abstract. It is the same question in miniature: am I building tools that earn back what they cost, or am I subsidizing a feature because it is impressive in a demo. Impressive and profitable are not the same test anymore.
Three separate threads this week point at the same shift: AI stopped being a line item you could ignore and became something that has to be managed like payroll or inventory. Here is what that actually looks like inside a business.
Token costs are now an operating expense, not a subscription fee. Several companies profiled this week burned through annual AI budgets in months because agentic AI behaves like variable labor, not flat-fee software. Every prompt consumes compute and electricity. The fix showing up at the smarter shops is not a blanket spending cap, it is a tiered budget with a clear path for an employee to ask for more once they show it is paying off.
Cheaper models are now good enough to change the math. Grok 4.6 landed at roughly sixty percent less cost per token than the leading alternative while matching it on benchmark tasks, and Ramp data showed one premium model capturing only six percent of enterprise token spend because businesses decided the extra performance was not worth the extra price. The lesson for anyone running a consulting practice or a small shop: stop assuming the most expensive model is the safest choice. Benchmark on your own tasks before you commit a client’s budget to a brand name.
Eliminating the entry-level seat has a bill that comes due later. A concept making the rounds this week, the tragedy of the cognitive commons, argues that checking AI output requires real expertise, and real expertise comes from years of doing the grunt work AI now eats first. Cut junior roles too aggressively and you lose the pipeline that trains the senior people who are supposed to catch AI’s mistakes ten years from now. That is a staffing decision posing as a technology decision, and it deserves the same scrutiny.
Anthropic’s own numbers may be the clearest signal yet. The company is reportedly targeting a two trillion dollar IPO on a run rate of one hundred to one hundred twenty billion dollars in annualized revenue, tripling for three straight years. Whatever you think of the valuation, the filing will be the first audited, public look at whether AI tokens are actually profitable at scale, or whether the entire industry is still being subsidized. That document is worth reading closely when it lands, not skimming for headlines.
· Rahm Emanuel put a number on a problem every AI infrastructure buildout runs into eventually: the country is short 500,000 to 600,000 electricians for data centers alone, plus 150,000 broadband workers. The bottleneck on AI right now is not chips, it is skilled trades. [All-In Podcast, August 13]
· Chai Discovery built what its founders call a neutral software factory for medicines, partnering with competing pharma giants at once instead of building its own drugs. It is a clean blueprint for any AI vendor selling into a regulated, IP-sensitive industry: stay neutral, segment the data, let the results sell themselves. [Lenny’s Podcast, August 11]
· A Databricks benchmark found Grok 4.6 completing tasks at 84 cents each versus $1.23 for the leading OpenAI model, with Elon Musk teasing a 4.7 release trained on proprietary SpaceX engineering data within weeks. Model choice is genuinely competitive again, not a two-horse race. [All-In Podcast, August 14]
· Jim Keyes, who ran both 7-Eleven and Blockbuster, made the case that fear, not lack of skill, is what actually sinks organizations facing disruption. Education is what interrupts that cycle. Worth remembering the next time a team resists a new tool out of instinct rather than reasoning. [Success Story with Scott D. Clary, August 11]
· A recurring line from Clay’s co-founder, shared in a company-wide writing policy that spread fast this week: if you generate a document from a short prompt and hand your readers the long version, you are disrespecting their time. Applies just as well to proposals and client emails as internal memos. [The AI Daily Brief, August 16]
Every business I talk to is running the same experiment right now, just with different budgets and different names for it. Some of that experimenting is generating real returns. Some of it is quietly eating margin nobody is watching closely enough to notice.
Where in your operation is AI still a demo you’re proud of, rather than a line item that pays for itself?
Hit reply and tell me what you found. I read every one.
Until next week,
Steve Parker
Founder, The Parker Group | AI Consultant, MBA, PMP
parkergroup.us | theparkergroup.substack.com
La Grange, Kentucky
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.