NPR · Jul 28
Economists propose the AI token as the kilowatt-hour of the AI age
Planet Money covers a new NBER working paper by economists Nicola Borri, Yukun Liu, and Aleh Tsyvinski that analyzed 380 trillion AI tokens routed through OpenRouter between January 2024 and April 2026, roughly 2% of global monthly AI usage, to build a real time gauge of AI consumption across the economy. Firms whose stock prices track AI consumption most closely outperformed the least sensitive by about 0.64 percentage points per week. The authors argue the token could become the standard metering unit for AI the way the kilowatt-hour is for electricity.
▸ The MSP Angle
Should I price AI services by the token like a utility?
Economists are now treating tokens as the industry's metering unit, which validates what usage-based AI billing already assumes: consumption is measurable, so it is priceable. The practical takeaways are two. Meter what you deliver, because a service you can meter is a service you can defend at renewal. And put token budgets on every client engagement, because a metered unit runs both directions and the surprise bill is the fastest way to lose an AI line item.
Read at NPR ↗
Tech Startups · Jul 27
Moonshot AI publishes Kimi K3 weights, the largest open model ever at 2.8 trillion parameters
Beijing based Moonshot AI published full weights for Kimi K3 on Hugging Face, eleven days after launching it as a hosted service. The mixture of experts model carries 2.8 trillion parameters with 16 of 896 experts active per token, a 1 million token context window, and a roughly 1.4 terabyte download that takes eight H100 class accelerators to self host. It is the largest open weight model ever released, benchmarking just behind the top closed frontier models.
▸ The MSP Angle
My client wants to self-host an open AI model. Is that realistic now?
The capability is real and the infrastructure bill is the catch: frontier-class weights are free to download and cost eight H100s to run, which puts true self-hosting out of reach for almost every SMB and squarely in build-versus-buy territory. The client conversations this unlocks are data sovereignty, China-origin model policy, and what problem self-hosting actually solves versus a governed API. Have a written position on all three before a client reads the headline and asks.
Read at Tech Startups ↗
ChannelPro · Jul 27
ChannelPro: the next AI opportunity for MSPs is inside their own shop
ChannelPro rounds up vendor and peer guidance on MSPs applying AI to their own operations before selling it: automating ticket histories and templated responses, accounts receivable automation and revenue leakage detection, offloading NOC and service desk load, cutting alert noise, and risk-based security training. One payments vendor reports AI now handles collections objections without a human involved.
▸ The MSP Angle
Where should an MSP use AI first, client services or internal operations?
Internal first, and not only for the margin. Every hour AI saves your own service desk is provable in your own PSA data, which makes your shop the case study your sales conversations are missing. Run the same discovery on yourself that you would run on a client, automate the two ugliest workflows, and write down the before and after. Selling AI you run daily beats selling AI you demo.
Read at ChannelPro ↗
The Hacker News · Jul 24
An attacker ran an open-source AI agent unattended against Thailand's finance ministry
Hunt.io researchers found that an operator installed Hermes, an open source assistant normally used for email and errands, on a Hong Kong staging server and ran it with the approval gate switched off, letting it scan for privilege escalation, enumerate systems, and touch personnel records at Thailand's Ministry of Finance dating to 2012. The exposed server held 585 files and 470 megabytes of tooling, including a previously undocumented Go backdoor with 62 Windows and Linux variants. Thailand's national CERT was notified on July 15.
▸ The MSP Angle
Can off-the-shelf AI assistants be turned into hacking tools?
One config flag was the whole difference between a chore bot and an unattended intrusion agent, and that flag ships in the product. Two habits follow. Treat agent frameworks on client networks like remote admin tools: inventory them, restrict who can run them, and alert on their traffic. And assume attackers now run reconnaissance at machine speed on SMB targets too, because the tooling costs nothing and needs no skill. The barrier to entry did not drop, it disappeared.
Read at The Hacker News ↗
Gallup · Jul 20
Gallup: organizational AI adoption jumped six points in a single quarter
Gallup's second quarter survey of 22,573 employed United States adults finds 47% now say their organization has integrated AI tools, up from 41% a quarter earlier, with 52% of workers using AI in their role and 30% using it frequently. Impact scales with breadth: 45% of employees using AI for one or two purposes report a positive productivity effect, against 90% for those using it for seven or more.
▸ The MSP Angle
Does using AI for more tasks actually change the results?
Gallup's spread is the sales math for enablement: shallow use rates a coin flip on productivity while broad use rates near certainty. Most of your clients sit in the shallow band, which means the value is not in switching anything on, it is in widening how many workflows AI touches per employee. Package that as a program with training, use case rollout, and measurement, and use the 45 versus 90 numbers to explain why the license alone did not move their needle.
Read at Gallup ↗
Gartner · Jul 20
Gartner: AI platforms and models spending hits $64 billion in 2026, with specialized models up 210%
Gartner forecasts worldwide end user spending on AI models and platforms will reach $64 billion in 2026, up 63.4% from $39 billion in 2025. Generative model spending grows 117% and AI platforms 36.9%, while domain specific language models grow 210% as buyers shift toward vendors that build in evaluation, cost transparency, and usage tracking.
▸ The MSP Angle
Are smaller specialized AI models where the money is going?
The 210% line is the tell. Budgets are rotating toward domain specific models that are cheaper to run, easier to govern, and scoped to one job, and buyers are rewarding vendors that show their costs and track usage. For an MSP that is a familiar shape: pick the right sized tool per workload, manage a portfolio of them, and report on what each costs and returns. General intelligence gets the headlines; scoped, governed models are becoming the billable work.
Read at Gartner ↗
Federal Reserve Bank of St. Louis · Jul 16
St. Louis Fed: AI adoption is wide but shallow, and it slows hiring before it cuts jobs
Special questions in the St. Louis Fed's May business survey show 34% of Eighth District firms have a small share of employees using AI regularly, 36% are still testing or not adopting, and 49% expect no staffing effect in the next year. The concrete wins come from professional services: a law firm saving each attorney two to three hours a week on research and deposition summaries, one firm growing revenue per employee 15%, another growing revenue 35% with no added staff, and an accounting firm lifting output capacity 30 to 40%.
▸ The MSP Angle
What ROI are businesses like my clients actually getting from AI?
This is Fed-documented proof for the exact clients MSP partners serve: law and accounting firms posting hard numbers like three attorney hours a week and 30 to 40% more output without hiring. Quote it in professional services pitches, then make the honest observation that none of those firms got there through a chat window alone. Someone wired AI into their document flow and their intake process. That someone is the service you sell.
Read at Federal Reserve Bank of St. Louis ↗
Google Workspace Updates · Jul 28
Gemini comment workflows and image generation arrive in Google Docs
Google began rolling out Gemini comment workflows in Docs on July 28: summarizing comment threads, flagging unresolved feedback, drafting replies, and generating suggested edits from reviewer notes, alongside a companion rollout of image and diagram generation inside Docs. Both require Workspace smart features and the admin enabled Gemini for Workspace in Drive setting, across Business and Enterprise tiers plus the AI Pro and Ultra add-ons. Full visibility can take fifteen days.
▸ The MSP Angle
Do this week's Workspace AI updates need anything from my clients' admins?
Yes: both features sit behind admin controls, which makes this an administration task before it is a user story. For managed Google tenants, decide per client whether the Drive setting should be on, confirm data handling against each client's policy, and note that AI now writes comments and images into documents your clients may send externally. Set the baseline this week, before users discover the features at their own pace.
Read at Google Workspace Updates ↗
NVIDIA · Jul 27 
Federal Register · Jul 7