At the @scsp_ai+ Expo last week, Arm’s Vince Jesaitis and Stephen Ozoigbo from our Government Affairs team led conversations on the future of AI and semiconductor growth. From scaling the physical AI stack to developing the workforce behind it, the discussions highlighted Arm’s role in enabling the performance, efficiency and scale needed for the next era of AI.

Meet the next-gen Razer Blade 18, our most powerful Blade ever, bringing desktop-class performance to a portable form factor: Powered by the Intel® Core™ Ultra 9 290HX PLUS processor and up to NVIDIA® GeForce RTX™ 5090 Laptop Graphics with 24GB VRAM, the Razer Blade 18 is built to power local AI development. From running models and testing agents to accelerating on-device inference, generation, and rapid iteration, it delivers the performance needed to build, create, and push workflows further. Available now exclusively at and RazerStores. #RazerBlade#

This is one of the more insightful things I've read this week: "By the late 1930s [the Germans] had developed the weapons that would in the main be used to fight the Second World War.....When the war came they tried to speed the process of development up, to win the war with the weapons of the 1950s. The result was a technical disaster: shortages of resources, constant political interference, the inherent difficulty of accelerating research work at the forefront of science, all meant that German forces got little in terms of performance from the new weapons to match the great expense of producing them. The Allies...stuck with the weapons of the late 1930s, and pushed them successfully to their limits, in most cases overtaking the performance of Germany's most conventional weaponry. When after the war they came to develop missiles, jets, advanced submarine technology, and a host of other vanguard equipment, they simply took German scientists and blueprints." Richard Overy, Why the Allies Won, 243.

I am the VP of Workforce Strategy at Meta and I built a spreadsheet called the Replacement Ratio that is, without exaggeration, the most elegant financial instrument in this building. Column A is headcount. Column B is quarterly CapEx allocation. Column C is what I call the Narrative Yield — how much each layoff announcement moves our price-to-earnings multiple. At Meta, cutting 8,000 people generates approximately 2.3x more shareholder value as a story than the $27 billion those people actually cost us. Like a controlled demolition where the dust cloud is worth more than the building ever was. I discovered this by accident in November 2022. We announced the first round on a Thursday. 11,000 people. The stock jumped 4% before market close. Our share price was $90 that week. I pulled up the actual savings — roughly $2.3 billion in annual compensation — and compared them to the market cap movement and the ratio was so disproportionate I thought I'd made an error. I had not made an error. I had discovered the Narrative Yield. The announcement IS the product. The terminations are just the input cost of producing it. Then Mark sent the second memo in March 2023. 10,000 more. "Flatter is faster," he wrote. "Leaner is better." "Keep technology the main thing." My team built talking points around each phrase. I remember testing "returning to a more optimal ratio of engineers to other roles" and watching three analysts independently upgrade the stock within 48 hours. Not because the ratio mattered. Because the sentence contained the word "optimal" and the word "ratio" and both of those words trigger the part of an analyst's brain that releases dopamine. We cut 21,000 people total. Our stock went from $90 to $600. Mark's net worth grew by approximately $170 billion. That is $9 million per fired employee. I calculated that number on a Tuesday afternoon and then went to get a coffee from the espresso bar in Building 40 that still operates at full capacity. The barista's name is Diego. He makes a very good cortado. He was not in any of the rounds. Our entire global payroll is $27 billion. Every engineer, every content moderator, every cafeteria worker who restocks the oat milk refrigerator in Building 21 next to the motivational poster that says EFFICIENCY IS CARING in Helvetica Bold, which was printed four days before we eliminated the internal print shop. All of them. $27 billion. Our CapEx guidance this year is $60 to $65 billion. Susan Li said it on the call in January — two weeks after we announced the latest round. The combined Big Four spend is $350 billion on AI infrastructure in 2025. Up from $165 billion just two years ago. If I fired every single employee tomorrow, all 72,000, the savings would cover maybe 42% of one year's data center buildout. The humans are a rounding error in the budget of machines that replace them. So what are the layoffs paying for? They are paying for the sentence. The one Susan Li reads on the earnings call: "These actions help us move more quickly while also helping to offset the substantial investments." That sentence is worth $40 billion in market cap. I know because I A/B tested the language with investor relations in March. We tested seven versions. Version C outperformed Version A by 340 basis points. Version C is the one with "actions" instead of "terminations." Version F used "workforce adjustments" and tested even higher but Legal flagged it as too close to the phrasing in the severance agreements. So we went with C. Turns out the market doesn't mind what you do. It minds what you call it. We call it a lot of things. "Flattening the org." "Removing redundancies." "Focusing our investments on our highest priorities." "Raising the bar on performance management." That last one was January 2025. Mark's memo. 3,600 people. He called them "lowest performers." The memo went out on January 14th. The earnings call announcing $60-65 billion in spending went out on January 29th. Fifteen days. My team scheduled both. The proximity is not accidental. You announce the human cost first so that when you announce the machine cost, the narrative is "disciplined" rather than "reckless." Sequencing is everything. We tested the reverse order once, hypothetically, in a simulation. The model predicted a 2.1% stock dip. Discipline first. Ambition second. Always. The performance framing was my suggestion. If you call them layoffs, it triggers severance obligations and unemployment benefits in thirty-seven states. If you call them performance-based terminations, it triggers nothing. Same people. Same desks cleared. Same badge deactivated at 5 AM before they woke up. Different word. Different $180 million in severance liability. I keep a legal pad in my desk where I track the savings per euphemism. "Performance management" saves approximately $50,000 per head in reduced severance. At 3,600 heads, that is $180 million. The cost of drafting the memo was forty minutes of Mark's time and sixteen hours of my team's time. That is approximately the best ROI in the history of corporate communications. Better than the Narrative Yield itself. Each phrase tests differently with different analyst cohorts. Growth-focused analysts respond to "investing in AI." Value analysts respond to "disciplined cost management." Same 8,000 people. Different sentence. Different $40 billion. The notification protocol is standardized now. Laptop access revoked at 5:47 AM Pacific. Badge deactivated at 5:48. Slack channels disappear at 5:49. Calendar cleared at 5:50. Personal email notification sent at 6:00. The thirteen-minute gap between systems going dark and the employee being told why is not cruelty. It is security protocol. We cannot have 3,600 people with simultaneous access to internal systems and knowledge that they have been terminated. The window for sabotage is too wide. So we close the window first and explain later. Some of them find out from the press release. Some of them find out because their phone loses work email at 5:47 and they check Twitter. I do not love this part. But I respect the engineering of it. Thirteen minutes. Clean. We announced the January cuts the same week Mark said "people will be more important than ever." My team wrote both statements. There is no contradiction if you understand that "people" and "headcount" are different financial instruments. People are the future. Headcount is the cost of having had a past. I keep a framed printout of both quotes side by side on my office wall. Not as irony. As a reminder that language is architecture. Meanwhile: we spent $77.86 billion buying back our own stock between 2022 and 2024. $27.96 billion. $19.77 billion. $30.13 billion. Each buyback inflates the share price. Each share price increase makes the layoff announcement look more justified in retrospect. The stock went up because we cut. We used the cash from cutting to buy back stock. The buyback made the stock go up more. The stock going up proved the cuts were correct. I mapped this loop on a whiteboard in January 2024 and one of our financial planning analysts took a photo of it and made it her laptop wallpaper. The total severance bill for 21,000 employees was approximately $2.5 billion. We spent 31 times that amount buying back stock. The humans cost less to remove than the stock cost to inflate. That is not a metaphor. That is the actual ratio. I have it in Column E. Reality Labs lost $60 billion between 2020 and 2024. Sixteen billion in 2023 alone. It was never subjected to the "Year of Efficiency." No one asked the metaverse division to be leaner or flatter or faster. The humans were asked to be efficient so the machines could be profligate. I did not design this asymmetry. I just maintain the spreadsheet that tracks it. The rehire pipeline is my favorite part. Half those roles reopen in Hyderabad and São Paulo within nine months at 31% of the loaded cost. Revenue per remaining employee went from $1.3 million in 2022 to $2.7 million in 2024. Each survivor now generates more than double what their predecessor generated. Not because they work harder. Because the denominator shrank and the numerator — AI-driven ad revenue — grew independently of human effort. We call it geographic rebalancing. The Workforce Transitions team keeps a Lucite tombstone on their shelf from the 2023 round, 11,000 MANAGED DEPARTURES etched in Helvetica, right next to a half-empty bottle of Clase Azul someone brought back from the offsite in Cabo where we planned the 2024 round. The same team is hosting a culture workshop next month called "Our People, Our Purpose." I wrote the talking points. Amazon is doing 30,000. Intel cut 21,000. Microsoft invented "voluntary departures" for 125,000 people, which is the most inspired euphemism since "rightsizing," because it implies the 125,000 chose this. Google cut 12,000 and called it a "moment of clarity." Salesforce eliminated 4,000 customer support roles and cited AI directly. Combined across the industry: 644,000 tech workers laid off since 2023. Combined CapEx on AI infrastructure: $350 billion this year alone. They spent seven to ten times more on GPUs than on severance for the humans those GPUs replaced. The layoffs are the press release for the spending. The spending is the excuse for the layoffs. It is a perpetual motion machine that runs on the difference between what a person costs and what their departure is worth. The free food budget for remaining employees is approximately $800 million per year. $10,000 to $12,000 per person. Artisanal pizza. Sushi bar. Pour-over coffee stations. The campus amenities operated without interruption during every round. Nobody asked the cafeteria to be efficient. I eat lunch there every day. It is very good. The oat milk is organic. Column D is the one I'm most proud of. It tracks average severance duration against local unemployment rates and cross-references media coverage density by market to optimize announcement timing for minimal news cycle disruption. January announcements get buried in earnings season. September announcements get lost in back-to-school cycles. I have mapped every dead zone in the American attention span and they are all on my calendar. January 14th — two weeks before Super Bowl coverage saturates every newsroom — was not an accident. The 3,600 number was calculated to stay below the threshold that triggers a WARN Act filing in California. 3,600 across twelve states. Below the threshold in each. That was also Column D. I presented the Replacement Ratio at our Q2 planning offsite last Tuesday. Someone from Legal asked if we'd modeled the human impact. I said yes. Column D. That's what Column D is. They promoted the spreadsheet to a standing dashboard. It refreshes hourly. Net income last year was $62.4 billion. Headcount is 72,000. The dashboard calculates revenue per head in real time. Every departure makes the number go up. Every departure makes the announcement worth more. Every announcement makes the stock go up. Every stock increase makes Mark $4.7 billion richer per percentage point. I named the Slack channel #narrative-yield#. It has 340 members. None of them are in Column A.

llama.cpp adds MTP for the Qwen3.6 family This is a significant milestone for the local AI ecosystem. The performance jump with these changes is massive and elevates local inference on commodity hardware further. Special thanks to Aman Gupta for leading this development!