If you think you're modeling skills are what's going to give you a long term career edge, you are wrong: > Thousands of hours spent developing the skill > Clueless nepo interns building crazy models > Tasks that would take 1hr+ are done in 2 mins You need to find a new edge

Hi, it's Darshita. I am a designer. And I'm pitching myself in public 'cause I'm done with a year full of anxiety. Everyone says design is dead and AI killed it. But the real parasite was hiding in plain sight. So here's my film, DESIGN IS DEAD. She's bold, fearless and very PINK 💖 Hiring? My DMs are open.

Big Boy rendered by Computer

Vivek Ramaswamy gave MAGA good vocabulary. He took their somewhat crude positions on meritocracy, immigration, executive power, anti-wokeism etc and wrapped them in the language of Yale and Harvard. He made the intellectual case for Trumpism better than Trump ever could. And now he is losing Ohio. A state Trump won by 11 points. There are many factors. Gas prices. Wars of choice. Trump's disdain for rural America. Vivek's rocky DOGE stint. His rebrand from national provocateur to state-level executive. But let us not pretend the obvious is not a factor. His race and his Hindu faith have been a constant undercurrent. In a fundamentally crude and unintellectual movement, the sophisticated vocabulary was borrowed. However, MAGA's instincts kicked in. MAGA used Vivek’s language when it was convenient. Now that he needs their votes, many have defaulted to their natural state. The movement that talks about meritocracy all day cannot bring itself to support the most meritocratic candidate on their side. Vivek is a billionaire. He neither needs nor deserves sympathy. But there is a deep irony here. He entered politics to expose the hypocrisy of the progressive left. He may end up exposing the hypocrisy of the populist right. Any wealthy, educated, brown-skinned person watching this race is coming the same conclusion: MAGA cannot look past your faith and skin color.

probably the best visual on how to run a personal / company operating system. here's my psychosis.

If you bought a new car in the last few years, go look in the trunk. There's a real chance the spare tire is gone and a can of goo is sitting where it used to be. Carmakers made that swap to squeeze out less than a tenth of a mile per gallon. In one year, AAA had to rescue more than 450,000 members who got a flat and had nothing to put on. Here's how lopsided that trade is. A spare weighs about 30 pounds, and the sealant kit that replaced it weighs about 4. Fuel economy rules grade carmakers on their whole fleet's average, and the Department of Energy's rule of thumb says every 100 pounds costs you about 1% in mileage. So ditching the spare bought around 0.3%. On a car that gets 30 mpg, that's 0.09 mpg. And the goo only works if the hole is in the tread and the nail is still stuck in the tire. Sidewall cut, blowout, or the nail popped out a mile back on the freeway? You're calling a tow truck. AAA says a refill can cost up to $300, and the stuff expires in 4 to 8 years. By 2017, 28% of new models came without a spare as standard. Pretty much everything in these photos went the same way. The carmaker saves a few pounds or a few bucks, and you end up paying for it later, usually on the side of the road. The touchscreen bugs me the most, because what it really costs you is your eyes on the road. A Swedish car magazine called Vi Bilägare had drivers do four simple things at 68 mph. Turn on the heated seat, bump the temp up two degrees and hit defrost. Find a radio station. Reset the trip meter. Dim the dash. In a 2005 Volvo with regular buttons, all of it took 10 seconds and 306 meters of road. In the worst touchscreen car they tested, the MG Marvel R, it took 44.6 seconds and 1,372 meters. That's most of a mile at highway speed, and every driver got practice time in every car before the test. So why'd carmakers do it? One sheet of glass replaces dozens of switches, and every one of those switches is its own part number, its own tooling, its own wiring and its own spot in a warehouse. They can also add features later with a software update. It got bad enough that Euro NCAP, Europe's big crash-safety rater, rewrote its 2026 tests so turn signals, wipers, hazards, horn and the SOS button only score if they're physical controls. The Audi Q7 scored 100% on those. The Xpeng X9 scored 20%. A crash-test agency is literally grading buttons back into cars now. The shifter has the darkest history. Back in 2016, Fiat Chrysler recalled 1.1 million Jeeps, Chargers and 300s over an electronic shifter that snapped back to center after every gear change, which made it hard to tell by feel what gear you were in. NHTSA tied it to hundreds of rollaways. Anton Yelchin's 2015 Grand Cherokee was part of that recall when it rolled down his driveway and killed him. Those old parts were doing a job nobody ever put on the window sticker. You could use them without looking. A real gear lever tells your hand it's in Park, and a knob clicks two degrees warmer while your eyes stay on the road. The average new car went for $50,562 last month, and it wants your eyes for stuff a knob handled by feel in 2005.

2026 is starting to look like 2008. Wall Street is buying insurance against an AI crash, and it's buying a lot of it. Most investors aren’t aware right now of what’s going on, but this matters for every tech and AI stock you own. The insurance is called a credit default swap. It pays out if a company can't pay its debts - the same contract that became famous in the 2008 financial crisis. In July, investors traded under $2B a week of this insurance on Oracle, Nvidia, Meta, Amazon, Alphabet, Broadcom and SpaceX. By September, it was over $9B a week. The reason is debt. AI companies have borrowed $466B so far this year to build out data centers (at this point last year, it was $101B). And the price of that protection is hitting records - insuring Meta's and Nvidia's debt has never cost more. For Oracle, the cost has more than doubled since January. CoreWeave is the extreme case. Insuring $10M of its debt costs $1.82M upfront, plus $125K every quarter. Some of its bonds already trade at distressed levels. The cracks are spreading. SoftBank fell 6% overnight as the cost of insuring its debt hit a 4-year high. In Korea, the biggest IPO on the tech exchange this year, a data center company, crashed 33% on its first day. Meanwhile, regular investors are doing the opposite. Nearly $10B poured into leveraged ETFs last week, the most on record, and $7B of it went into a fund that bets 3x on chip stocks. But I don't think this is a bubble. These AI companies are growing earnings 30% a year, and the market trades at 19 times earnings, nowhere near the 25 times we saw before the 2000 crash. But when the smart money starts paying up for protection, sentiment is without doubt shifting. Higher rates and higher energy prices make all that borrowing a lot more expensive. So keep cash on the sidelines. You're going to get to buy a lot of these AI stocks back 25-30% lower. Be patient, and you'll probably do very well over the next couple of years…

J.P. Morgan Asset Management’s base case projects Bitcoin to outperform U.S. stocks, bonds and gold over the next decade. Wall Street is doing the math.

Sound on!

$IREN: Close Collaboration with Nvidia to Develop Open Source Security for AI Clouds Just as Linux became the dominant OS for x86 CPUs, the Nvidia Open Source stack will become the dominant platform for AI Clouds. It’s fully vertically integrated software-firmware-hardware and part of Nvidia’s expanding moat. The IREN announcement at GTC will be advancements in AI Cloud Security for Agents on Nvidia’s stack.

$SPCX SpaceX Q3 FY26: • Revenue +161% Y/Y to $14.6B ($1.6B beat). • Operating margin 19% (+17pp Y/Y). • CapEx $19.6B (including $16.4B for SI). • GAAP EPS $0.35 ($0.19 beat).

RETENTION is becoming one of the BIGGEST problems for Meta's Muse. According to data shared by Morgan Stanley, 73% of Muse users are active on Day 1, but by Day 14, that number collapses to just 23%. And what's even more concerning is that Instagram, Google, and Facebook are all showing significantly stronger retention over the same period. This tells us that while Meta might be attracting users initially, it's having a much harder time convincing those users to actually stick around. And as someone who used Muse HEAVILY when it first came out, I can tell you that compared to Grok, Muse is a complete JOKE. The experience simply isn't on the same level, and I think this highlights one of the BIGGEST challenges facing the entire AI agent industry right now. Think about it. What exactly is stopping someone from switching from Muse to Grok, ChatGPT, or another AI agent tomorrow? Absolutely NOTHING. There are very few switching costs and unless these companies can build products that become deeply integrated into people's daily routines, users are going to continue jumping between whichever platform offers the best experience. And here's why this matters so much for Meta. We're talking about a company spending billions of dollars building AI infrastructure, training increasingly sophisticated models, and trying to establish itself as one of the leaders in artificial intelligence. But none of that matters nearly as much if consumers aren't actually using the products consistently. AI infrastructure is incredibly expensive and without sustained engagement, it's going to become increasingly difficult to justify the economics behind these investments. Getting people to download an application is the EASY part. Convincing them to come back every single day is where the real competitive advantage gets built. Now, I wouldn't completely write Meta off because the company still has one of the BIGGEST distribution advantages in the world through Instagram, WhatsApp, and Facebook. It also has the financial resources to keep improving its models and products. But distribution alone doesn't guarantee retention, especially when consumers have so many alternatives available to them.

Absolutely incredible… "Wedgewood Polls" just admitted they are a fake polling firm and made up their polls. Their fake polls have been posted in both the NYT and Nate Silver’s averages.

Huge thanks to @claudeai for featuring my wildfire analysis prototype!🙏 So cool to see it alongside all these amazing projects

$SPCX $IREN $NBIS $CRWV $BE The line from the SpaceX call every neocloud holder should save. Musk: "the power cooling electrical equipment infrastructure is somewhat of a limiting factor." SpaceX is Nvidia-exclusive, is bringing on ~200,000 GB300s a month, and plans to keep that pace next year with Vera Rubin. The thing it's worried about is power, cooling and switchgear. Musk said he's been spending "often four days a week" in Memphis to get the cluster online. What SpaceX is doing about it: • Aiming for an incremental 20 GW of power and cooling next year against ~10 GW of incremental compute, so that power and cooling is "far in excess of what is needed to turn the GPUs on." In August he put it as "far more power, cooling and electrical equipment than we have GPUs" • Building that infrastructure at a rate Musk says is "vastly faster than anyone else and may at some point be faster than everyone else combined" • Still pushing the compute target up. Year-end went from "over 2 GW" in August to 2.5-3 GW on this call, and for 2027: "I would be surprised if we're brought on less than 13 gigawatts of compute by the end of next year" If a company with ~$90B of cash and securities and Musk on site four days a week is bottlenecked on power, think about what an energized, grid-connected MW is worth to everyone else. Who already has it: • IREN: ~5 GW of power under grid agreements, the 1.4 GW Sweetwater 1 substation energized since May, and the 2 GW Sweetwater hub conditionally included as base load in ERCOT's Batch Zero. It agreed to buy Nostrum for 490 MW in Spain and announced an 800 MW campus in South Australia, with ~0.5 GW of IT load targeted for 2027 and 2026 capacity "largely sold out." Co-CEO Daniel Roberts: "Every megawatt we build is scarce" • Nebius: 0.8-1 GW connected by year-end, a 5 GW contracted power target, two US campuses of up to 1.2 GW each in Missouri and Pennsylvania, and more than 1 GW deployed per year from 2027. CFO: "as fast as we bring capacity online, we can sell it" • CoreWeave: over 1.5 GW active, past 1.85 GW by year-end, and contracted power up from 3.7 GW to 4.2 GW in six weeks. The FT reported nearly all of its 2027 capacity is sold. Brannin McBee: "If you gave me 100 megs of powered shell today, I'd sell it in 18 seconds" • Bloom: on-site power that doesn't have to wait years for a grid interconnect. Nebius alone has 328 MW of Bloom fuel cells coming online this year Now put SpaceX's price on it. Above $50 per watt is $50M per MW per year. At that rate 100 MW is $5B a year and 1 GW is $50B. The neoclouds sign long-term at $20-25M/MW, so the same GW is $20-25B a year. Either way the revenue goes to whoever can turn the power on first. And look at what the market pays for that power today: • IREN at ~$13B EV is ~$2.6M per MW of power under grid agreements • CoreWeave at ~$80B EV including debt is ~$19M per MW of contracted power • Nebius at ~$47B is under $10M per MW of its 5 GW contracted power target Building out a MW costs a lot more than that, mostly in GPUs. But GPUs can be ordered, and the power is the part SpaceX says it can't get fast enough. That's also why SpaceX would rather overbuild power than run short. At $50M per MW a year, 1 GW of GPUs sitting idle waiting on a substation is ~$4B of lost revenue every month. It's hard even for SpaceX. Nameplate went from 1.0 GW to 1.4 GW to 1.7 GW over the last three quarter-ends, and the new plan needs ~2.5 GW a quarter in 2027. Last week the market sold every company with energized power and signed contracts as if AI demand had broken. This week one of the biggest GPU buyers in the world, and now one of the biggest sellers of compute, said its limit is how fast it can power and cool buildings.

The value of Terafab’s output will place it in the top 20 of country GDP in a single building

Elon Musk just SAVED the neocloud trade. Elon “Our belief is that the value of compute per watt will increase over time, not decrease. And the reason that for that, I think it should be obvious to people really is that as you improve the AI software, as the models get better and better, they're able to you get more useful intelligence per watt. And the value of that compute is proportionate to the useful intelligence per watt. And that's why you see things like even old computers like H100s renting for more than they rented when they were new. And I think that's likely to continue. I think we're going to face a shortage of superintelligence compute”.

AI, now called Super Intelligence, has overtaken Starlink to become SpaceX's biggest business by revenue. SpaceX ->SIX

$IREN: @ShanuMathew93 Debunks Overly Aggressive 2027 Power Buildout Thank you @ShanuMathew93 for debunking the first fact given that breaks the whole model: "27 GW cited from ERCOT is forecast growth in total Texas peak electricity demand over 2026–28". First this is the forecasted demand not forecasted generation from a April 2025 ERCOT forecast (1)! It was stated in the same paper (1) and Matthew pointed out that only "55.4% of officer-letter projects scheduled for 2024 had energized". ERCOT only met roughly half of what was demanded. Yes, some of it is double booked in other grids but you cannot use the 27GW ERCOT requests to be new power added onto ERCOT grid. Second, 27GW is for all industries and there's alot of other industries in TX. Only 11.4GW is for forecasted for datacenters over 2026-2028 (1). Thirdly, @bubbleboi is looking at swapping "~1.5 Blackwell racks for one Rubin rack at the same power". Don't you understand that if you don't build capacity for the new GPUs, then you have to build capacity for the old GPUs. There are no available DC capacity right now. It's like musical chairs: 10 people, 10 chairs, when you add 5 people, you have to take 5 people out. You either build capacity for the blackwells or you toss them. Now consider what would break if we would just upgrade and toss out the old ones: the whole thesis of high profitability rest on GPU lifetimes. Tossing out a B200? Hell no lol, B200 came out in 2024, and you want to toss in 2027? A 3 year GPU life time? You going to find some F-grade DC overseas with 50% uptime and 6 months of shipping and reracking time? Great use of B200s that you just paid off and broke even on. You going to toss A100s? Do datacenters that run aircooled A100s at lower rack density run liquid cooled VRs at high rack densities? If financing is an issue, why not just keep renting out A100s that are all paid off. CRWV is doing it right now. Fourthly, that's not how SB6 works. SB6 forces everyone to have onsite backup so there is a buffer for critical loads like hospitals, residential, etc. If you use that buffer to optimize the number of datacenters, then that goes back to square one: having no buffer. Fifth, @bubbleboi states that crypto miners will add another ~5GW. Those 5GWs are part of 27GW of demand request stated cleaning in the chart as crypto. I guarantee you BTC is not valuable enough for there to be 5GW of BTC miners in Texas alone right now. Sixth, @bubbleboi talks about using recips. Go ask $ORCL and $NBIS how all their recip sites now re-designed to be $BE sites are going.

BofA: Data center additions ramp hard from 13 GW in 2025 to ~56 GW by 2030E: The industry is expected to add 212 GW of cumulative incremental capacity between 2026-2030 Nvidia still the core of the stack but custom silicon like is scaling aggressively alongside AMD and others $NVIDIA $GOOGL $META $MSFT $AMZN $AMD