# AI capex — X 热门讨论 (2026-09-24 07:23 UTC)
## @NinNinnin0306 (NinNin) · 09-24 03:22 · ♥83 ↻0 💬95 AI can generate images of the world. But Physical AI needs to understand the real one
A robot doesn’t just need text or images
It needs to understand streets, buildings, objects, distances, layouts and how physical environments actually change
That’s why spatial data is becoming an important part of the Physical AI stack
Vangrid is building a human collected spatial data layer designed to provide real-time ground truth for world models and autonomous systems
The interesting part is the pipeline:
- Capture real world environments - Process data with edge computed privacy - Build spatial information - Add cryptographic provenance - Make it accessible through a Spatial API
Vangrid is essentially trying to turn a distributed network of contributors into a zero capex sensor swarm for the physical world
The bigger question isn’t just how much data AI can collect
It’s whether that data can actually help machines understand the world they operate in
That’s the part I’m watching with Vangrid
@vangrid_io > 引用 @NinNinnin0306: What if every smartphone could become a sensor for Physical AI ?
AI models can learn from billions of tokens
But robots need something different
They need to understand streets, buildings, objects, distances and how physical environments actually look
That makes real world spatial data an important piece of the Physical AI stack
This is where Vangrid gets interesting
Vangrid is building a decentralized spatial data layer around human-collected data, with the goal of providing continuous ground truth for world models and autonomous systems
The pipeline is pretty interesting:
- Capture real world environments - Process data at the edge - Reconstruct spatial information - Attach cryptographic provenance - Make the data accessible through a Spatial API
There’s also a bounty layer where operators can submit 3D captures, while payments are handled through USDC escrow on Base
To me, the bigger idea is simple:
Physical AI needs a way to continuously learn what the real world looks like
Vangrid is trying to build that data layer
@vangrid_io https://x.com/NinNinnin0306/status/2102961892834890163
## @glocalinvestor (Arvind Srinivas) · 09-23 12:38 · ♥41 ↻11 💬0 JPM issues their updated view on Memory market
Covered major memory debate & shared contrasting bullish and bearish scenarios along with its own view on the debate. Make sure to follow for more.
My Aim: Don't tell what to do. Help you decide what to do.. You can find my read at the end.
#1: Is AI Capex sustainable? - Bullish: AI business remains lucrative & hyperscalers find it not challenging to access more capital - Bearish: Slow business adoption resulting in low investments from hypescalers - JPM View: Robust hyperscaler investment. No slowdown in memory atleast until FY28.
#2: Is memory budget in AI capex sustainable? - Bullish: Indispensable. Memory burden justified to power compute & token gen. Ltd tech breakthrough. - Bearish: Memory share might move to pre-AI <10% levels. Optics, SRAM or non-HBM soln imminent. - JPM View: Leaning towards bullish scenarios. But expects milder ASP growth. Big enough Compute demand to bear reduced per compute memory
#3: Is HBM de-spec a danger? - Bullish: Content downgrade is to power installation growth. HBM mix in $DRAM continues to grow. - Bearish: Evidence of peaking HBM performance & fading economics. Demand reroute to conv. DRAM - JPM View: Negative to total HBM demand but not thesis breaking. HBM vs DRAM Bit demand shows HBM growing faster. Training demand for ~24 mths.
#4 Is LTA critical? - Bullish: Building structurally healthy customer-vendor relationship. Greater visibility for both - Bearish: Will fall apart during down-cycle. Caps incremental ceiling by limiting ASP increase. - JPM View: Bullish for memory. Accounts 70% of capacity implying CSPs desperation to secure memory. Reduced cyclicality of the industry
#5 Is China threat real? - Bullish: Capacity is real but still has tech gap which limits them from accessing entire market - Bearish: Caps the cycle the old fashioned way: Supply glut resulting in demand destruction. - JPM View: China is a structural threat and has to be closely monitored. Limited near-term risk due to technology gap.
My read: Memory is surely near term (FY27/28) story. Jury very much out whether the structural demand is sustainable or not. For now, I would keep accumulating Samsung / $SKHY / Kioxia / $MU and $SNDK
Key Catalysts: - Concrete LTA disclosure - Shareholder return - Memory S/D sufficiency - AI capex https://x.com/glocalinvestor/status/2102739361645129736
## @trevornoren (Trevor Noren) · 09-23 20:58 · ♥34 ↻4 💬3 Goldman: "Almost half of S&P 500 growth in EPS in 2026 comes from AI investment. The largest US hyperscaler companies are on track to spend $800b on capital expenditures this year, an increase of 94% over 2025. That money is flowing through the earnings of chipmakers, tech hardware suppliers, industrial firms, and utilities. There are second-order effects too. The boom has lifted capital markets activity and supported consumer spending through rising household wealth. However, both consensus and Goldman Sachs analyst forecasts show hyperscaler capex growing at a slower rate in coming years. In the meantime, the hyperscalers’ equipment carries depreciation charges that keep climbing as spending growth slows. This will further dampen the boost of AI investment spending to S&P 500 earnings growth."
Whether it's enterprise adoption, unit economics, or data center buildout challenges, I see myriad threats to market expectations about how soon and at what scale AI will deliver ROI. But even beyond the viability of AI ROI, AI-trade outperformance faces structural headwinds and it's not just comps. To quote my recent report on "The AI Trade" (https://t.co/wQQNniS2kj):
"Comps are a significant concern for US equities over the next 12 months. UBS has estimated that hyperscaler AI CAPEX will increase roughly 75% this year, 25% next year, and only 6% in 2028. That rate-of-change slowdown will certainly hit the semiconductor industry the hardest. However, we also believe market participants have neglected the impact circular spending is having on hyperscaler growth and the trickle-down impact to companies across sectors. As the rate of increase decreases, we expect earnings growth to broadly slow. And that’s before considering how everything we’ve laid out in this report could lead to spending that falls below estimates.
The comp headwind will be paired with another headwind: a spike in equity supply. In the run-up to SpaceX’s IPO, the surge in new equity supply grabbed headlines. That concern has since fallen off given SpaceX’s IPO did not have a measurable downside impact on the market. However, we wouldn’t take a short-term pullback in concern as evidence of long-term irrelevance. YTD, IPOs have raised $137b, putting this year on pace to far exceed the previous record—$156b in 2021. Added to that is the new supply coming from tech giants. In June, Alphabet alone raised ~$85b in the largest public stock sale in history. New supply is set to only increase over the next year. More mega-IPOs are coming. Meanwhile, 90- to 180-day lockup periods will expire, and selling by insiders, employees, and pre-IPO investors usually translates to more new supply than the IPO itself. Investors shouldn’t underestimate how that new supply could exacerbate a market downturn if AI sentiment dips and drags on inflows."
Chart link: https://t.co/AK0P0ll6Xk https://x.com/trevornoren/status/2102865303298588829