Jason Shouldice

[Jason Shouldice @ShouldiceJason]

We just fired our call center's biggest bottleneck: waiting for a manager to notice.

Meet Sarah. She's an AI floor manager, and she runs our call center floor. Not a dashboard. Not a report you read on Friday. She runs the floor.

Here's what she does, every single day, with no one telling her to:

🎧 She listens to every call. Not 2%. Not a random sample. Every call, both sides, scored against our script within about 2 minutes of hanging up.

🔊 She checks every rep's line every 60 seconds. Muffled mic, TV in the background, Zoom audio leaking into calls, bad internet, dead air. She hears it before the customer complains.

📹 She runs the morning meeting. She opens her own Zoom, greets every rep by name as they join, and gives a 30-minute training in her own voice: yesterday's numbers, the script line by line, and a real booked call broken down into 6 lessons.

🔒 She enforces the standard. Miss the opener 4 times in 5 calls? Sarah waits until your call ends, puts you on Training for 10 minutes, and won't let you back on the phones until you type the opener back to her, word for word.

⚡ She coaches live. Tips in your DMs. Call of the moment posted to the team. 10 perfect openers in a row? She calls you out by name in front of everyone.

⏱️ She doesn't let anyone hide. 25-minute "15-minute break"? She noticed. Sitting on the after-call screen? She noticed. Logged in late? Noted, and in the report.

🛟 She can't break the floor. Max half the team on Training at once. Every lock unlocks itself. If any tool fails, she releases everyone and alerts us.

📊 She reports. Scoreboard every hour. A personal scorecard for every rep at 8 PM: their numbers, their best call, and the 3 things to fix tomorrow.

The old model: 1 manager, 20 reps, 3 calls reviewed a day, feedback a week late.

The new model: every call, every minute, every rep, coached in real time.

Call centers don't have a people problem. They have a feedback-speed problem. Sarah fixes it.

This is the future of the floor, and it's already running ours.

[Jason Shouldice @ShouldiceJason]

🚨 ELON MUSK JUST PUT A $500 BILLION–$1 TRILLION SPACEX REVENUE RUN RATE ON THE TABLE FOR THE END OF 2027.
⠀
Read that again.
⠀
Five hundred BILLION.
To ONE TRILLION.
Annualized.
By the end of NEXT YEAR.
⠀
That is Musk’s forecast. It is not booked revenue. It is not a guarantee.
⠀
But before you dismiss it—or blindly believe it—you need to understand the industrial machine taking shape underneath that number.
⠀
Because the history makes this story considerably more explosive.
⠀
Rewind to 2024.
⠀
On October 13, SpaceX caught a returning Super Heavy booster with giant mechanical arms.
⠀
A skyscraper-sized rocket booster came back out of the sky and a tower CAUGHT IT.
⠀
That actually happened.
⠀
Meanwhile, in Memphis, xAI built the original Colossus system with 100,000 NVIDIA Hopper GPUs in 122 days, according to NVIDIA.
⠀
One hundred thousand GPUs.
One hundred twenty-two days.
⠀
At the end of 2024, Starlink reported serving more than 4.6 million customers.
⠀
Those are the historical anchors.
⠀
A rocket catch.
A 100,000-GPU supercomputer.
Millions of satellite internet customers.
⠀
Now move forward.
⠀
During 2025 alone, Starlink says it added MORE THAN 4.6 MILLION new active customers.
⠀
In one year, it added a customer count comparable to the entire base it reported at the end of 2024.
⠀
Then came September 2026.
⠀
Starship Flight 14 reached orbit and deployed 26 Starlink V3 satellites.
⠀
Starlink’s published design specification for V3:
⠀
ONE TERABIT PER SECOND of downlink capacity.
PER SATELLITE.
⠀
Twenty-six satellites.
Approximately 26 terabits per second of aggregate design downlink capacity.
⠀
That is not a promise that every customer gets a terabit connection. Geography, spectrum, utilization, gateways, and terminals still matter.
⠀
But the size of the infrastructure step is real.
⠀
And now SpaceX wants to attempt something even bigger:
⠀
Catch the booster.
Catch the ship.
Recover both stages of the orbital launch system.
⠀
Then work toward flying them again.
And again.
And again.
⠀
Recovery is the first hurdle.
Reliable, cheap, rapid reuse is the economic prize.
⠀
If they reach it, the implications extend into every business that benefits from putting more hardware into orbit.
⠀
More communications capacity.
More payload.
More opportunities to build services that are currently too expensive.
⠀
And the compute business is already producing serious numbers.
⠀
SpaceX reported approximately $14.6 BILLION in Q3 revenue.
⠀
Approximately $7.6 BILLION came from its Super Intelligence segment.
⠀
Approximately $5.2 BILLION came from connectivity.
⠀
AI was already the largest revenue segment.
⠀
Let that land.
⠀
The company famous for rockets now has an AI segment generating more quarterly revenue than its connectivity segment.
⠀
And this is where the story gets absolutely wild.
⠀
Anthropic publicly confirmed an agreement to use all the compute capacity at Colossus 1.
⠀
Its announcement described access to more than 300 megawatts and over 220,000 NVIDIA GPUs.
⠀
Google’s agreement was disclosed in an SEC filing:
⠀
Approximately 110,000 NVIDIA GPUs and supporting infrastructure.
⠀
Full-rate payments of $920 MILLION PER MONTH.
⠀
That annualizes to $11.04 BILLION.
⠀
From one customer agreement.
⠀
There are termination rights. The full headline contract duration is not an unconditional guarantee of every future payment.
⠀
But the commercial relationship is real.
⠀
These companies are paying for access to physical machines.
⠀
And the next fight is over how fast those machines can be connected to POWER.
⠀
Because a GPU sitting in a warehouse does not run your business.
⠀
It needs electricity.
Cooling.
Networking.
Switchgear.
Transformers.
Buildings.
Commissioning.
⠀
You can order the silicon.
⠀
You still have to build the industrial system around it.
⠀
That is why one of the most consequential parts of Musk’s remarks involved something almost nobody casually following AI talks about:
⠀
TURBINE BLADES AND VANES.
⠀
Yes.
⠀
The race toward increasingly powerful AI now runs through precision metal casting.
⠀
SpaceX says it is applying its manufacturing capabilities to components for terrestrial power turbines.
⠀
The strategy is straightforward:
⠀
Find the bottleneck.
Build capacity around the bottleneck.
Get the power online.
Turn on the compute.
⠀
The shortage itself is not imaginary.
⠀
GE Vernova reported 116 GIGAWATTS of gas-power equipment backlog and slot reservations in Q2.
⠀
That does not mean SpaceX has solved turbine manufacturing.
⠀
It means the constraint it is attacking is substantial.
⠀
Musk says internal production could create a 12–18-month advantage in obtaining those components.
⠀
That advantage still needs to be demonstrated in qualified parts, manufacturing yields, and commissioned power systems.
⠀
But consider what even a meaningful fraction of that advantage could be worth when customers are willing to spend billions for earlier access to compute.
⠀
Now to the number that explains the revenue headline.
⠀
Musk says he expects approximately 13 GIGAWATTS of compute by the end of 2027.
⠀
He also described a smaller possibility of exceeding 20 GW.
⠀
Targets.
Not installed capacity today.
⠀
Management discussed monetization around $30–$50 per watt annually, with current pricing described as above the top of that range.
⠀
Here is the simple arithmetic:
⠀
1 GIGAWATT = 1 BILLION WATTS.
⠀
13 GW × $30 per watt annually = $390 BILLION.
⠀
13 GW × $50 = $650 BILLION.
⠀
20 GW × $50 = $1 TRILLION.
⠀
THAT is the mathematical foundation beneath the headline.
⠀
The multiplication is easy.
⠀
Delivering every assumption inside it is brutally hard.
⠀
You need the equipment.
You need working facilities.
You need customers.
You need utilization.
You need pricing.
You need financing.
You need operating reliability.
⠀
And you cannot count the same capacity twice.
⠀
A GPU training your own model cannot simultaneously be rented at full capacity to someone else.
⠀
Internal training needs to earn its return through the products it creates.
⠀
Still—
⠀
You can now see why this is a serious industrial story.
⠀
The revenue target is connected to an attempt to build and monetize an extraordinary amount of physical computing infrastructure.
⠀
For perspective:
⠀
A continuous 13 GW electrical load consumes approximately 114 TERAWATT-HOURS per year.
⠀
Twenty GW consumes approximately 175 TWh.
⠀
The scale is staggering.
⠀
Then there is Terafab.
⠀
The stated target is a functional test chip around May or June 2027.
⠀
Volume production around Q3 2028.
⠀
The ultimate ambition is chips representing approximately ONE TERAWATT of compute power annually.
⠀
A working test chip is not proof of competitive AI processors at commercial yields.
⠀
A factory announcement is not production.
Production is not yield.
Yield is not profitability.
⠀
Those distinctions matter enormously.
⠀
But so does the ambition:
⠀
Build more of the chip supply.
Build more of the power infrastructure.
Build more of the compute.
Build the AI products.
Expand global connectivity.
Reduce the cost of moving hardware into orbit.
⠀
My read is that they are trying to control more of the physical dependencies underneath the AI economy.
⠀
And now both SpaceX and Tesla have formed special committees to explore a potential merger.
⠀
No agreed terms.
No guaranteed transaction.
No certainty that shareholders ultimately approve anything.
⠀
But the strategic overlap is obvious enough to take seriously:
⠀
AI software.
Chip design.
Robotics.
Manufacturing.
Connectivity.
Launch infrastructure.
⠀
The giant question is whether combining the companies creates enough additional value to justify the price and governance consequences.
⠀
That remains unresolved.
⠀
Here is what I find most extraordinary about the timeline:
⠀
2024:
A first Super Heavy tower catch.
An original 100,000-GPU Colossus build.
More than 4.6 million Starlink customers reported by year-end.
⠀
2025:
More than 4.6 million NEW Starlink customers added.
⠀
2026:
Starship reaches orbit and deploys V3 satellites.
SpaceX reports approximately $14.6 billion in quarterly revenue.
Its AI segment contributes approximately $7.6 billion.
Major technology companies buy access to its compute.
⠀
2027 TARGET:
Approximately 13 GW of compute.
A possible larger upside case.
A $500 billion–$1 trillion annual revenue run rate by year-end.
⠀
2028 TARGET:
Terafab volume chip production.
⠀
The distance between those dates is tiny.
⠀
The distance between those capabilities is enormous.
⠀
There are plenty of ways this can fall short.
⠀
Launch delays.
Manufacturing problems.
Power bottlenecks.
Lower compute prices.
Customer cancellations.
Expensive internal AI workloads.
Capital requirements.
Execution failures.
⠀
SpaceX reported approximately $19.6 BILLION in Q3 capital expenditure.
⠀
The ambition requires enormous spending before all the hoped-for returns arrive.
⠀
So nobody should confuse an extraordinary forecast with a completed outcome.
⠀
But dismissing the entire story also means ignoring infrastructure already built, customers already connected, hardware already launched, and commercial agreements already signed.
⠀
That is what makes this so fascinating.
⠀
The future being proposed is extreme.
⠀
The starting point is already substantial.
⠀
If even a meaningful portion of this plan works, the competitive consequences could be enormous.
⠀
More available compute.
More capable agents.
More businesses able to deploy them.
More pressure on companies whose operations still depend on slow, repetitive computer work.
⠀
As someone building AI systems inside a business, THAT is the part I am watching.
⠀
What useful work can the next generation of infrastructure make economically possible?
⠀
How many workflows?
How many customer conversations?
How much software?
How much analysis?
How much execution?
⠀
The milestones to watch are concrete:
⠀
Rockets recovered and reflown.
Satellites commissioned.
Megawatts energized.
GPUs serving workloads.
Customers paying.
Cash returns catching up with capital spending.
⠀
That is where this extraordinary forecast either becomes an extraordinary business—
⠀
or collides with reality.
⠀
Either way, the next 24 months are going to be WILD. 🚀

[Jason Shouldice @ShouldiceJason]

@TheSolSaverX @thsottiaux Right?

[Jason Shouldice @ShouldiceJason]

@TheSolSaverX @thsottiaux He sends a reset several times a month surely your on the upside of wining

[Jason Shouldice @ShouldiceJason]

@aaronjmars @brian_armstrong @coinbase I know right wtf lol

[Jason Shouldice @ShouldiceJason]

@thsottiaux thanks big daddy

[Jason Shouldice @ShouldiceJason]

@pxue Love the path

[Jason Shouldice @ShouldiceJason]

🚨🤯 𝗜 𝗗𝗢𝗡’𝗧 𝗧𝗛𝗜𝗡𝗞 𝗣𝗘𝗢𝗣𝗟𝗘 𝗙𝗨𝗟𝗟𝗬 𝗖𝗢𝗠𝗣𝗥𝗘𝗛𝗘𝗡𝗗 𝗛𝗢𝗪 𝗙𝗔𝗦𝗧 𝗔𝗜 𝗛𝗔𝗥𝗗𝗪𝗔𝗥𝗘 𝗜𝗦 𝗠𝗢𝗩𝗜𝗡𝗚.

So let’s put ACTUAL NUMBERS on it.

Forget the hype.

Look at what happened in FOUR YEARS.

━━━━━━━━━━━━━━━━━━

🟢 𝟮𝟬𝟮𝟮 — 𝗡𝗩𝗜𝗗𝗜𝗔 H100

This thing was an absolute MONSTER when it arrived.

One H100 SXM:

🧠 80 GB HBM3 memory
⚡ 3.35 TB/sec memory bandwidth
🔗 900 GB/sec NVLink
🔥 ~4 PFLOPS sparse FP8 Tensor compute
⚡ Up to 700 watts

Companies started spending BILLIONS building H100 clusters.

This was the hardware that helped ignite the modern generative-AI explosion.

Now fast-forward just FOUR YEARS.

━━━━━━━━━━━━━━━━━━

🔴 𝟮𝟬𝟮𝟲 — 𝗡𝗩𝗜𝗗𝗜𝗔 RUBIN

ONE Rubin GPU:

🧠 288 GB HBM4

⚡ Up to 22 TB/sec memory bandwidth

🔗 Up to 3.6 TB/sec NVLink

🔥 50 PFLOPS sparse NVFP4 inference

🔥 17.5 PFLOPS dense FP8/FP6

🔥 4 PFLOPS dense FP16/BF16

🤯 336 BILLION transistors

Compare just the memory systems:

H100: 80 GB
Rubin: 288 GB

That’s roughly 3.6X the memory capacity.

H100: 3.35 TB/sec
Rubin: up to 22 TB/sec

That’s roughly 6.6X the peak memory bandwidth.

H100 NVLink: 0.9 TB/sec
Rubin NVLink: up to 3.6 TB/sec

That’s roughly 4X the interconnect bandwidth.

IN FOUR YEARS.

But that’s STILL not the crazy part.

━━━━���━━━━━━━━━━━━━

🏭 𝗧𝗛𝗘 𝗚𝗣𝗨 𝗜𝗦𝗡’𝗧 𝗘𝗩𝗘𝗡 𝗧𝗛𝗘 𝗖𝗢𝗠𝗣𝗨𝗧𝗘𝗥 𝗔𝗡𝗬𝗠𝗢𝗥𝗘.

NVIDIA is building AI at RACK SCALE.

One Vera Rubin NVL72 contains:

🔥 72 Rubin GPUs

🧠 36 Vera CPUs

🧠 20.7 TERABYTES of HBM4

⚡ ~1.4 PETABYTES/SECOND of HBM bandwidth

🔥 3.6 EXAFLOPS of sparse NVFP4 inference

🔗 ~216 TB/sec aggregate NVLink bandwidth

🧠 Up to 54 TERABYTES of CPU memory

💻 3,168 custom NVIDIA CPU cores

𝗜𝗡. 𝗢𝗡𝗘. 𝗥𝗔𝗖𝗞.

🤯🤯🤯

Think about 1.4 PETABYTES/SECOND.

That’s roughly:

1,400 TERABYTES of data moving every SECOND.

The machine can move an amount of data equivalent to roughly 1,400 one-terabyte SSDs EVERY SECOND through its aggregate HBM memory system.

━━━━━━━━━━━━━━━━━━

🚨 𝗡𝗢𝗪 𝗛𝗘𝗥𝗘’𝗦 𝗧𝗛𝗘 𝗣𝗔𝗥𝗧 𝗜 𝗖𝗔𝗡’𝗧 𝗦𝗧𝗢𝗣 𝗧𝗛𝗜𝗡𝗞𝗜𝗡𝗚 𝗔𝗕𝗢𝗨𝗧.

H100 → Rubin:

FOUR YEARS.

So what happens during the NEXT FOUR YEARS?

Nobody knows.

So let’s NOT pretend we know the specifications of a 2030 GPU.

Instead, let’s make a thought experiment.

Suppose the next four years give us ONLY another 10X improvement in economically useful AI inference.

Call this:

🔮 𝗧𝗛𝗘 𝟮𝟬𝟯��� 𝗦𝗖𝗘𝗡𝗔𝗥𝗜𝗢.

If today’s Rubin rack represents “1X”…

Imagine a future equivalent system delivering:

🔥 ~36 EXAFLOPS worth of useful low-precision inference capacity

⚡ ~14 PETABYTES/SECOND-class effective memory movement

🤖 ~10X the useful agent throughput

💰 ~1/10 the compute cost per unit of useful work

Again:

THOSE ARE NOT ANNOUNCED NVIDIA SPECS.

They’re a thought experiment showing what ANOTHER 10X improvement would mean.

And this is where things get absolutely insane.

━━━━━━━━━━━━━━━━━━

🤖 𝗪𝗛𝗔𝗧 𝗗𝗢𝗘𝗦 𝟭𝟬𝗫 𝗠𝗢𝗥𝗘 𝗜𝗡𝗙𝗘𝗥𝗘𝗡𝗖𝗘 𝗔𝗖𝗧𝗨𝗔𝗟𝗟𝗬 𝗕𝗨𝗬 𝗨𝗦?

People think:

“ChatGPT answers 10X faster.”

NO.

That’s probably one of the LEAST interesting consequences.

Imagine today you can economically throw 10 serious AI agents at a problem.

Tomorrow you throw 100 at it.

Instead of ONE coding agent:

🤖 20 agents design solutions.

🤖 30 agents write different pieces of software.

🤖 10 agents independently attempt alternative implementations.

🤖 10 agents run tests continuously.

🤖 10 agents hunt bugs.

🤖 5 agents attack security.

🤖 5 agents optimize performance.

🤖 5 agents simulate users.

🤖 4 agents review everybody else’s work.

🤖 1 frontier model acts as the chief engineer.

ALL SIMULTANEOUSLY.

And they don’t sleep.

They don’t get tired.

They can run thousands of experiments.

They can throw away failed solutions and immediately try again.

neously.

The scarce resource stops
nce COLLAPSE.

And if that happens…

[Jason Shouldice @ShouldiceJason]

@LinkedInHelp @Arun_Kumar451 Hey Ray I appreciate your help! Kindly check for a DM from me. Thanks.