Funding the AI bonanza

Investment Analysis Clubs Macro Economic Trends and Risks
1

The investments in AI hardware and software are absolutely mind-boggling. Some of this is being funded from the operating earnings of AI hyperscalers like Alphabet, Amazon and Meta. But there’s plenty coming from outside.

https://www.wsj.com/finance/investing/global-stocks-markets-dow-news-06-08-2026-aac7c547?mod=hp_lead_pos2

Wall Street Is Rushing to Fund the AI Bonanza in Every Conceivable Way

From giant debt deals to IPOs, tech companies keep raking in investor cash

By Sam Goldfarb, The Wall Street Journal, June 8, 2026

  • Tech companies are raising billions through equity and bond sales to fund the artificial-intelligence build-out, with Alphabet announcing an $85 billion equity raise.

  • AI hyperscalers have issued $159 billion in bonds globally this year, up from $17 billion in 2024, to finance data centers.

  • Despite concerns about overbuild risk, investor demand for AI-related investments remains strong, with tech stocks up 31% this quarter.

Funding rounds and IPOs raising 11-figure sums. Blockbuster bond sales spanning three continents. The casual announcement of an $85 billion equity raise.

Such is life on Wall Street at the dawn of the artificial-intelligence build-out. Tech companies are hungry for cash to invest in data centers, and investors are forking it over through all possible means, in all parts of the globe—a flurry of fundraising that has mostly supported markets by powering technological advances, even as it tests their ability to absorb it all…

image
image1107×853 23.6 KB

Plenty of investors still think that the AI build-out could get messy, with companies overspending and Wall Street eventually weeding out the losers. In the meantime, increased equity issuance could pressure stocks by diluting shareholders—one possible factor behind last week’s tech selloff. …

Spending on data centers and other AI infrastructure by just four big tech companies this year is expected to total more than $670 billion—a larger investment as a share of the economy than even the railroad expansion of the 1850s…

At this point, even some skeptical investors say there is little point in betting against the AI boom when it appears to be just getting started…[end quote]

Just getting started? ROFL! The tech hyperscalers already represent 41% of the S&P500 index! Tech stocks up 31% this quarter. How is this “just getting started” in a way that will profit an investor who is jumping in with fresh cash?

Goldman Sachs released an updated research note covering the massive acceleration in capital expenditures. They project that the four primary hyperscalers—Alphabet, Microsoft, Amazon, and Meta —will spend a combined $5.3 trillion in capital expenditures between fiscal 2025 and 2030.

If you expand the lens over the next five years to include the broader tech industry ecosystem (adding Apple, Oracle, Tesla/xAI, private equity infrastructure funds, and global utility grid upgrades), Goldman estimates total baseline AI infrastructure spending will hit roughly $7.6 trillion . To put that into perspective, $5.3 trillion is larger than the entire annual gross domestic product (GDP) of economic powerhouses like Japan or India.

When even the skeptics throw in the towel and say they are just getting started it’s time to start the stopwatch on the bubble pop. How long will it take for profits from actual end users (who are not part of the AI build-out ecosystem) to pay for all this?

image
image890×494 37.6 KB

Wendy

5 Likes
2

@WendyBG

As I use Claude to code, it thinks over and over again, giving me breaks to do simple surfing. I have not read this, but the headline tells a good story. I will read it later tonight.

There is a great deal of economic promise and progress in this development, but it may drastically outdate earth bound DC. At the other end the PC and Mac Mini are going AI, which drastically reduces the cost of AI to the end user. Again outdating earth bound DC.

PS, I got the privacy terms a few minutes ago from Anthropic. They state we do not have paid ads in our service. Gemini does. For Gemini this makes sense. People use Google to find products. The differences are much more problematic for Chat.

1 Like
3

Here’s a YT link. 32 minutes.
The interview is superb.
But a lot of folks will judge it OT.

:rocket:
ralph

3 Likes
4

I will never have enough time in the day. Tonight is limited. Tomorrow is completely booked.

5

Elon Musk has a way of turning conventional methods upside down and AI hardware is no exception.

Why tesla’s AI5 chip is so cost effective

GoogleAI:

Tesla’s AI5 chip achieves immense cost-effectiveness—costing under 10% of comparable Nvidia enterprise hardware—primarily through radical simplicity, vertical integration, and purpose-built efficiency. By designing silicon exclusively for their own localized inference workflows, Tesla eliminated expensive, generic components found in traditional GPUs. [1, 2, 3, 4, 5, 6]

Key Drivers of AI5 Cost-Effectiveness

  • Radical Simplicity: Because Tesla is their own and only customer, the chip doesn’t need to accommodate generic third-party workloads. Tesla deleted legacy GPUs, image signal processors, and redundant general-purpose computing blocks, yielding a streamlined, half-reticle die size that is cheaper to manufacture. [1, 2, 3, 4]

  • Software-Hardware Co-Design: Instead of building a brute-force chip and adapting software to it, Tesla’s software requirements dictate the hardware architecture. The chip is heavily optimized for low-precision mixed-precision math (e.g., INT4, INT2, FP8), accelerating bottlenecks like native softmax operations which previously required multiple emulation steps on AI4. [1, 2, 3]

  • Advanced Packaging & Memory Integration: The AI5 uses a sophisticated packaging strategy that integrates memory dies right next to the central System-on-Chip (SoC). Shortened interconnects provide massive bandwidth and inference efficiency without inflating total silicon area or power demands. [1]

  • Dual-Fab Manufacturing Scale: Manufactured domestically in the US across facilities from TSMC (in Arizona) and Samsung (in Texas), Tesla aims to overproduce these chips. Sharing the silicon across millions of electric vehicles, Optimus humanoid robots, and data centers drives down unit costs via extreme economies of scale.

While the world is spending gazillions on nebulous AI targets Tesla is creating a relatively inexpensive chip to power RoboTaxis and the Optimus humanoid robot. The expected financial outcome could not be more divergent.

The Captain

5 Likes
6

Captain what beget what?

Yes, both Microsoft (with its custom Copilot+ PC chips and Cobalt CPUs) and Apple (with the M-series in the Mac Mini) are pursuing similar strategies to Tesla’s AI5 chip. All three are shifting to highly integrated, purpose-built silicon to maximize efficiency and control their own ecosystems. [1]

The parallels between these companies highlight a major shift in computing:

  1. Tailored Silicon vs. Generic Hardware
  • Tesla AI5: Removes legacy components and third-party compatibility needs to focus entirely on localized AI inference.
  • Apple Silicon: M-series chips integrate the CPU, GPU, unified memory, and specialized Neural Engine onto a single System-on-Chip (SoC), stripping away generic motherboard bottlenecks.
  • Microsoft Copilot+ PCs: Microsoft is partnering with chipmakers (Qualcomm, Intel, AMD) to mandate dedicated Neural Processing Units (NPUs) in Windows PCs, explicitly designed to run local AI tasks rather than relying on heavy, off-the-shelf GPUs. [2, 3, 4, 5]
  1. Software-Hardware Co-design
  • Tesla AI5: Architecture is built entirely around Tesla’s specific software needs and low-precision math.
  • Apple: Integrates macOS deeply with hardware, allowing applications to tap directly into the unified memory and Neural Engine for massive performance-per-watt gains.
  • Microsoft: Developed the Windows Copilot Runtime and APIs first, then worked directly with hardware manufacturers to ensure processors have the exact instruction sets to accelerate these features locally.
  1. Vertical Integration and Control
  • Tesla: Controls the entire stack from the proprietary FSD (Full Self-Driving) neural networks down to the AI5 silicon inside every vehicle and robot.
  • Apple: Controls both macOS software and the proprietary architecture of the Mac Mini, resulting in highly optimized power efficiency.
  • Microsoft: While Microsoft does not manufacture its own consumer silicon at scale (partnering instead with external foundries), it dictates the exact hardware specifications required for its AI-first Windows ecosystem.

The primary difference is that Tesla builds hardware exclusively for its own closed ecosystem of vehicles and robots, whereas Apple and Microsoft build computing platforms for millions of diverse consumers and developers.

If you are interested, I can:

  • Detail the specific AI performance specs of the Apple M4/M-series
  • Explain how Microsoft’s Copilot+ NPU requirements differ from traditional Intel/AMD processors
  • Compare the power efficiency of these consumer chips against Enterprise Nvidia hardware [6]

Let me know how you’d like to explore this further.

AI responses may include mistakes.

[1] https://www.linkedin.com/posts/kit-yu-27753534_custom-silicon-efforts-like-maia-detailed-activity-7431700557419065344-mPXs

[2] https://aimultiple.com/edge-ai-chips

[3] https://riscv-europe.org/summit/2025/media/proceedings/2025-05-13-RISC-V-Summit-Europe-16h30-VALERIO-slides.pdf

[4] https://www.articsledge.com/post/neural-engine

[5] https://www.microcenter.com/site/mc-news/article/this-week-in-ai-jan-2-2026.aspx

[6] https://fudzilla.com/amds-ryzen-ai-400-desktop-apus-land-with-zen-5-and-rdna-3-5/

3 Likes
7

Yep. I still think it’s a risky investment, but you discount Elon at your peril.

Did you see that Elon filed for a permit to have 5,000, steering wheel-free, CyberCabs roaming the streets of Las Vegas?

intercst

1 Like
8

Didn’t say CyberCabs, didn’t say steering wheel-free:

26-05015 Tesla Robotaxi LLC AVNC Application Notice

Seems likely they’ve finally got FSD to the point where they can run driverless in the Loop.

2 Likes
9

Article said that CyberCab operation is included under the permit. Elon has pointed out many times that 90% of taxi hails have 2 passengers or less. Thus, the fleet mix will look something like 90% CyberCab, and 10% Model Y Robotaxi, or other larger vehicles.

The whole point is cost efficiency, an area where Waymo is far behind.

intercst

3 Likes
10

Ah yes, many of us have “plans” and one of us has much grander plans.

2 Likes
11

Sure - but that doesn’t mean that they’re planning on deploying CyberCabs, any more than they have in Austin. It just means they’re allowed. But they could just continue using modified Model Y’s, just like they do in Austin.

How do you know Waymo is far behind? Sure, it is expensive (per car) to convert existing vehicles into AV’s by adding all the sensors and stuff after the fact. It is also expensive (per car) to build an entire production line and then only manufacture a few hundred of that vehicle. Yes, once Tesla scales up to needing 10K or more vehicles per year, the cost of that production line per car will come down. But once Waymo has approached that scale, they’ll undoubtedly also pivot to mass production of purpose-built AV’s instead of what they’re doing now.

All Tesla’s done is put the capital into the production line a year or more earlier than they need it. That doesn’t provide cost efficiency, and may end up being more inefficient than the approach of waiting until you need the cars to build the production line.

Plus, it may turn out (and IMHO is likely to turn out) that vehicle cost is a relatively modest part of the TaaS cost profile, and the whole thing will turn more on operational costs and efficiency and the infrastructure to support them. And we have no information about whether Tesla or Waymo is “far behind” on that.

8 Likes
12

You really going to take that bet?

13

That Tesla is going to deploy 5,000 steering-wheel free CyberCabs in Nevada? Absolutely.

I will wager 10 Internet Points against all comers that Tesla does not have 5,000 steering-wheel free CyberCabs deployed in Nevada within the twelve (12) month timeframe covered by that permit.

3 Likes
14

It is always interesting. He’s out of the political crap. He has the personal chops to do things.

15

A short list of things Musk didn’t invent

  • Electricity
  • ASICs
  • EVs
  • Batteries
  • Solar panels
  • Cars

BTW, I don’t know about the Microsoft chip but the Apple M1 is not about AI inference.

All investing is risky and one does it at one’s peril when not managing OPM.

The CyberCab tsunami is coming! :slightly_smiling_face:

The Captain

16

Captain,

The memory aspect can carry the LLM. All three of them are doing this on the desktop, so to speak, not literally in Tesla’s case.

Yes, the memory bottleneck is the universal truth across all three chips. While each company brands its AI hardware differently, the true performance enabler for all of them is massive memory bandwidth and shared architectures, not just raw compute cores.

Here is how this rule applies to Tesla, Microsoft, and Apple.

Apple M1: The Blueprint

  • The Claim: Apple heavily markets its Neural Engine. [1]
  • The Reality: The Apple M1 chip succeeded because its Unified Memory Architecture (UMA) puts the CPU, GPU, and Neural Engine on a single package. They all share one pool of high-bandwidth memory. [2, 3, 4, 5, 6]
  • Why it matters: AI models are massive. Moving data from standard RAM to a separate graphics card creates a massive speed bottleneck. Apple bypassed this by letting the AI cores read data instantly from the main memory pool. [7, 8, 9]

Microsoft (Qualcomm Snapdragon X & Intel/AMD): The Adaptor [10]

  • The Claim: Microsoft pushes “Copilot+ PCs” based on NPU (Neural Processing Unit) performance metrics. [11, 12, 13]
  • The Reality: To meet Microsoft’s strict local AI requirements, silicon partners had to copy Apple’s homework. Chips like the Qualcomm Snapdragon X Elite utilize wide, high-bandwidth LPDDR5x memory channels soldered close to the chip.
  • Why it matters: An NPU with high compute power is completely useless if it has to wait for slow, traditional PC system memory to feed it data. Microsoft’s ecosystem is shifting away from upgradeable RAM sticks toward soldered, ultra-fast unified layouts specifically to feed these AI workloads. [14, 15]

Tesla (FSD Computer & Dojo): The Extreme Case

  • The Claim: Tesla designs custom AI chips (like the Full Self-Driving computer and Dojo) optimized for neural networks. [16, 17, 18, 19]
  • The Reality: Tesla’s architecture is the ultimate proof that memory is everything. In their car computers, the specialized AI accelerator cores are surrounded by massive pools of ultra-fast graphics memory (GDDR6) or system memory with direct, high-bandwidth paths.
  • Why it matters: For autonomous driving, latency kills. Tesla’s chips do not just need to process images; they need to stream gigabytes of video data from multiple cameras into the AI model simultaneously. If the memory architecture cannot feed those video frames to the AI processing cores instantly, the car cannot react in real time. [20]

Architecture Comparison

Feature [21, 22, 23, 24, 25] Apple M1 Series Microsoft (Copilot+ / Snapdragon) Tesla FSD / Dojo
Primary AI Marketing Neural Engine NPU (Copilot+) FSD Computer / Dojo Nodes
Real Secret Sauce Unified Memory Architecture High-bandwidth soldered LPDDR5x Ultra-low latency SRAM & custom high-speed buses
Primary Bottleneck Solved Moving data between CPU and GPU System RAM latency in traditional PCs Streaming massive video/sensor data instantly

If you are interested, we can look deeper into this. I can:

  • Break down the exact memory bandwidth speeds (GB/s) of these three systems
  • Explain the difference between SRAM, DRAM, and UMA in AI processing [26, 27, 28]

Let me know which technical direction you would like to explore next!

[1] https://www.volico.com

[2] https://hacarus.com

[3] https://www.macstories.net

[4] https://www.electropages.com

[5] https://www.linkedin.com

[6] https://www.everythingdevops.dev

[7] https://www.instagram.com

[8] https://www.tlciscreative.com

[9] https://www.mycomputerworks.co.uk

[10] https://www.tweaktown.com

[11] https://spectrum.ieee.org

[12] https://www.cio.com

[13] https://www.computerworld.com

[14] https://news.futunn.com

[15] https://medium.com

[16] https://tspasemiconductor.substack.com

[17] https://www.extremetech.com

[18] https://www.klover.ai

[19] https://www.levelfields.ai

[20] https://www.veltyx.de

[21] https://medium.com

[22] https://www.linkedin.com

[23] https://en.eeworld.com.cn

[24] https://daringfireball.net

[25] https://github.com

[26] https://semiengineering.com

[27] https://blog.roundhillinvestments.com

[28] https://www.linkedin.com

17

Meanwhile, the cybercab count in Austin is… 20. And it’s still not the entire city.

1 Like