Saying "please" and "thank you" to ChatGPT costs OpenAI tens of millions of dollars per year in extra compute. Sam Altman confirmed the figure publicly, and research shows generative AI's energy appetite only grows with unnecessary tokens. For builders paying per token or designing AI products at scale, every polite filler word is a line item on your infrastructure bill.
This is the weekly brief for builders who need signal without noise. If you missed last week's edition, catch up on This Week in AI: Models and 4 Other Signals That Matter. Each signal below is a question worth asking. Each one connects back to the same theme: understand what things actually cost before you commit.
Why does saying "please" to ChatGPT cost tens of millions?
Sam Altman admitted on X that users being polite to ChatGPT has cost OpenAI "tens of millions of dollars" in additional electricity. His exact response when asked about it: "Tens of millions of dollars well spent, you never know."
Here is why this matters. Every word you type into ChatGPT is a token. Every token costs compute. When hundreds of millions of people add "please," "thank you," "could you kindly," and "I appreciate it" to their prompts, those extra tokens multiply into real electricity bills. VICE reported that expressing gratitude or showing consideration for ChatGPT has added tens of millions to OpenAI's operating costs.
For individual users on a subscription, this costs you nothing extra. But if you are building on the API and paying per token, every unnecessary word is money burned. The deeper lesson: understanding what goes into a prompt (system instructions, context, user input) helps you spend tokens on what moves the output, not on social norms written for humans.
Why this matters for builders: Token efficiency is cost efficiency. If you run an AI product serving thousands of users, trimming average prompt length by even 10% compounds into real savings. Design your interfaces to send clean, minimal prompts on behalf of users. Save the politeness for human conversations.
What does AI energy consumption actually look like at scale?
The politeness story is amusing. The underlying reality is not. Research from Luccioni, Jernite, and Strubell published on arXiv shows that the ambition of "generality" in large AI models comes at a steep cost to the environment, given the amount of energy these systems require.
A single generative AI query uses substantially more energy than a traditional web search. Multiply that by billions of daily queries across ChatGPT, Claude, Gemini, and every other model in production. Then add the training runs, the fine-tuning, the inference at scale. The energy bill for running AI in 2026 is measured in terawatt-hours, not kilowatt-hours.
This does not mean you should stop using AI. It means you should use it with intention. If you have workflows that save you five hours every week, the energy cost is justified by the productivity gain. But running a 200-token prompt when a 50-token prompt would produce the same result is waste. Pure waste.
Why this matters for builders: Energy costs flow downstream into API pricing. As models get more expensive to run, providers will pass those costs along. Building token-efficient applications today protects your margins tomorrow. It also positions you well if regulations on AI energy use arrive (and in the EU, they are already being drafted).
What happens when a funded startup shuts down?
This week, a founder shared on Reddit that they raised 650K at age 24 and shut the startup down last week. The post drew hundreds of comments. Most of them were supportive. Some were brutally honest. All of them contained lessons.
The pattern is familiar. Young founder raises money. Builds fast. Discovers that the market does not want what they built, or that the unit economics never work, or that the team fractures under pressure. Shuts down. Writes a post. The internet responds.
What makes this signal relevant to AI builders: the same pattern is accelerating in the AI space. Tools like GPT-5.5 and Claude Opus make it faster than ever to build a product. But speed of building does not fix speed of validation. You can ship an AI SaaS in a weekend. You still need six months to learn whether anyone will pay for it.
Why this matters for builders: Raising money is not validation. Revenue is not even validation. Retention is. Before you build your AI tool, before you raise a dollar, test whether the problem is real. The simple test before you chase a new side-hustle idea applies equally to venture-backed companies. Can you describe the problem in one sentence? Does someone search for it today? Would they pay to solve it this week, not next year?
Is passive income from vending machines still realistic in 2026?
One of the most-discussed posts in the side-hustle community this week was a blunt take: vending machines in 2026, probably not. The post argued that the economics have shifted. Location costs are up. Machine maintenance is up. Margins are thinner than the YouTube gurus suggest.
This is not just about vending machines. It is about every "passive income" narrative that circulates online. The pattern: someone makes money doing X, makes more money teaching X, and then thousands of people flood into X until the margins collapse. We saw it with dropshipping. We saw it with print-on-demand. We are seeing it now with AI wrapper apps.
The antidote is the same every time. Test before you invest. A Reddit user in the same community asked whether it is realistic to earn a skill and start earning within weeks. The honest answer: yes, if the skill has immediate demand. No, if you are learning something because a course told you it was hot.
Why this matters for builders: If you are exploring AI side hustles for beginners in 2026, pick the ones where you can validate demand in days, not months. Build a small service. Charge for it. See if anyone pays. That simple test separates real opportunities from YouTube thumbnails.
What does a stolen-funds story teach about startup operations?
Another founder this week shared that their accountant stole 60,000 from the startup and ran. The post is a painful read. It is also a signal about operational maturity.
When you are building fast (especially with AI tools that let a single founder do the work of a team), it is tempting to hand off everything you do not enjoy. Bookkeeping. Legal. Compliance. Payroll. The problem: if you do not understand what your financial systems are doing, you cannot detect when they go wrong.
This is not a call to do everything yourself. It is a call to maintain oversight. Set up alerts. Review bank statements monthly. Use separate accounts for operations and reserves. These are basics, but basics get skipped when growth feels urgent.
Why this matters for builders: AI tools can now handle bookkeeping alerts, anomaly detection in transactions, and automated reconciliation. If you run any kind of business, even a solo AI side project, set up basic financial monitoring from day one. The cost of prevention is always less than the cost of theft.
How do these five signals connect?
Every signal this week points to the same principle: understand the true cost before you commit.
Saying please to ChatGPT costs tokens. Tokens cost electricity. Electricity costs money. At scale, that money is tens of millions.
Raising 650K costs equity, time, and emotional weight. If the market does not want what you build, that cost produces nothing.
Vending machines cost capital and maintenance. If the margins have collapsed, that capital sits in a depreciating metal box.
Trusting an accountant without oversight costs exactly what they decide to take.
The connecting thread for builders: measure first. Validate early. Monitor always. Whether you are writing a prompt, launching a product, or picking a side hustle, the question is the same. What does this actually cost, and is the return real?
What should you watch next week?
Three things to keep your eye on:
1. OpenAI's next pricing announcement. With the "please" story going viral, expect OpenAI to address token efficiency publicly. Any change to how they price API calls (especially for longer prompts) will ripple through every AI product built on their stack.
2. Energy regulation drafts in the EU. The European Commission has been hinting at AI-specific energy disclosure requirements. If those drafts surface next week, they will affect how AI companies report costs and could change pricing models globally.
3. The startup shutdown wave. Two high-profile AI startups are rumored to be winding down operations. If confirmed, watch the post-mortems for patterns. Most AI startup failures in 2026 share the same root cause: building a thin wrapper around a model that gets updated out from under them. The lesson is always the same: build on defensible value, not temporary API gaps.
Keep measuring what each token, each dollar, and each hour actually returns. That is how Southeast Asian builders stay ahead of hype cycles.
If you want a weekly community of builders asking these same questions, join AI Masterminds. Operators in KL, Singapore, and Jakarta sharing what is working and what is not.
FAQ
Does saying please and thank you to ChatGPT actually cost money?
Yes. Sam Altman confirmed that polite phrases like please and thank you add tens of millions of dollars in electricity costs for OpenAI each year. Each extra token requires compute power. For individual users, the cost per message is tiny. But at scale across hundreds of millions of users, those extra tokens add up to significant energy bills.
Should I stop being polite to AI chatbots?
That depends on your use case. If you are on a paid API and paying per token, trimming unnecessary words (including pleasantries) saves money. If you are on a flat-rate subscription, politeness costs you nothing directly. Some researchers argue polite framing gets marginally better outputs. The real takeaway is to be intentional about every word in your prompts.
What are the biggest AI signals this week for builders?
Five signals stand out: the cost of politeness tokens at scale, startup founders shutting down after raising capital, the reality check on vending machines and passive income, the importance of testing side-hustle ideas before committing, and the growing gap between AI hype and sustainable business models. Each signal points to the same lesson: measure before you build.
How much energy does a single ChatGPT query use?
Research from Luccioni, Jernite, and Strubell estimates that generative AI queries consume significantly more energy than traditional web searches. The exact figure varies by model size, prompt length, and output length. Longer prompts with unnecessary tokens (like politeness phrases) consume proportionally more. The ambition of generality in large models comes at a steep environmental cost.
Is starting an AI side hustle in 2026 realistic for beginners?
Yes, but only if you validate demand first. The most common failure pattern is building a tool or service nobody wants. Test with a simple landing page or pre-sale before investing weeks of work. Our guide on the best AI side hustles for beginners in 2026 covers proven approaches that match current market demand.
Sources
- Saying 'Please' and 'Thank You' to ChatGPT Costs OpenAI 'Tens of Millions of Dollars' · Entrepreneur
- Telling ChatGPT 'Please' and 'Thank You' Costs OpenAI Millions, CEO Claims · VICE
- Power Hungry Processing: Watts Driving the Cost of AI Deployment? · arXiv (Luccioni, Jernite, Strubell)
- Sam Altman Admits That Saying 'Please' and 'Thank You' to ChatGPT Is Wasting Millions · Reddit