Key Takeaway: What three years inside a sovereign wealth fund during the GlobalFoundries acquisition teaches about scale, category selection, and the GPU road not taken.
I spent three years inside one of the world's largest sovereign wealth funds. The decision-making process looks nothing like what outsiders imagine. From the outside, it looks like a pool of capital making rational investment decisions. From the inside, it's a political organization where investment theses must survive multiple committees with conflicting mandates, and the timeline between identifying an opportunity and deploying capital can exceed eighteen months. The talent is extraordinary -- some of the sharpest people I've ever worked with. The frustration is watching that talent operate inside a structure that was designed for consensus, not speed.
The approval chain
The investment team identifies and develops the thesis. The internal investment committee reviews against strategic priorities — first political filter. Strong returns with no alignment to capability-building get deprioritized. The board and sovereign stakeholders provide final approval, where dynamics shift to statecraft.
Abu Dhabi's investment strategy was explicitly about post-oil diversification. Semiconductor manufacturing and aerospace received massive capital not because returns were superior to index funds, but because they created domestic capabilities. When the best investment also builds capability, the system works. When they diverge, politics wins — the capital is a nation's, not a pension fund's.
Why speed is structurally impossible
$500M to $5B checks require extensive diligence and regulatory approvals. Confidentiality constraints slow information flow. Consensus requirements extend timelines across stakeholders who don't share the same information or priorities. The talent is world-class. The frustration is the gap between what the people can do and how fast the institution lets them do it.
The GlobalFoundries Story
In 2009, Abu Dhabi acquired AMD's manufacturing arm and created GlobalFoundries. I was at ATIC (Advanced Technology Investment Company, Abu Dhabi's tech investment arm, later merged into Mubadala) during the transaction, managing the portfolio integration.
The thesis was reasonable: Abu Dhabi understood oil wealth had a finite horizon. AMD needed billions to keep fabs competitive. TSMC was pulling away — a well-capitalized alternative should attract customers.
But the thesis rested on an assumption that turned out to be the single most consequential bet in modern semiconductor history: that unbundling was the future.
The unbundling bet
The entire semiconductor industry in 2009 was moving toward separation. AMD had just split its design and manufacturing arms -- that's what created GlobalFoundries. The logic was clean: design companies design chips, foundries manufacture them, and the interface between them is standardized. Specialization wins. This was the "fabless" revolution that had already produced Qualcomm, Broadcom, and TSMC's dominance.
NVIDIA made the opposite bet. Jensen Huang was building a vertically integrated stack -- hardware, software (CUDA), developer ecosystem, and increasingly, the applications themselves. In 2009, this looked like a niche strategy for a graphics card company. CUDA had launched two years earlier to lukewarm reception. The machine learning community was small and academic, years away from the deep learning explosion that AlexNet would trigger in 2012.
The semiconductor world measured itself against Intel's x86 monopoly because that's where the $200 billion server market lived. Nobody in the room -- not AMD, not Abu Dhabi, not the investment bankers -- framed GPUs as a strategic computing platform. They were graphics cards. NVIDIA was a $10 billion company. Today: over $2 trillion.
Why bundling won
The unbundling thesis assumed that the value in semiconductors was in the physical manufacturing -- the atoms. NVIDIA proved the value was in the integration -- the system. CUDA made NVIDIA GPUs programmable for general compute. The developer ecosystem created lock-in that no amount of manufacturing scale could match. When deep learning arrived and needed massively parallel processing, NVIDIA didn't just have the hardware -- they had the software stack, the tooling, the documentation, and a decade of developer relationships.
GlobalFoundries could manufacture chips. But NVIDIA owned the full stack from silicon to software, which meant customers weren't buying chips -- they were buying capabilities. That's the difference between a commodity and a platform.
Was the unbundling bet wrong at the time? It was the consensus view, and consensus was earned. The fabless model had created enormous value. What nobody anticipated was that the next paradigm shift -- AI compute -- would reward deep vertical integration over specialized separation. The technical signals were there if you knew where to look: Huang was already evangelizing general-purpose GPU computing, and the parallel processing architecture that made GPUs good at rendering pixels turned out to be exactly what neural networks needed. But semiconductor strategy in 2009 was fought on the CPU battlefield, and the fringe signals came from a handful of researchers running neural nets on gaming GPUs.
The lesson for bundling decisions
This isn't just a semiconductor story. The bundling question recurs everywhere: should you own the full stack or specialize? The answer depends on where value accrues in the next paradigm, not the current one.
When the interface between layers is well-defined and stable, unbundling wins -- specialization creates efficiency. When the interface is shifting or when the value comes from deep integration across layers, bundling wins -- the system is worth more than the sum of its parts. Apple understood this with iPhone. NVIDIA understood it with AI compute. The mortgage industry -- fragmented into eight separate layers with standardized handoffs -- is a textbook unbundled system. The question is whether AI will shift value to whoever controls the integrated data-and-decision layer, rewarding rebundling the same way it rewarded NVIDIA.
Sovereign wealth funds, by design, optimize for consensus and strategic alignment. The GPU opportunity required contrarian conviction that the institutional structure couldn't produce. The lesson isn't that GlobalFoundries should have predicted AI. It's that category selection in capital-intensive technology is the bet that matters most, and the signals for category shifts come from the fringe, not the mainstream.
The yield problem. TSMC's dominance is accumulated process knowledge that can't be purchased. At a new node, TSMC hits 80% yield within months; a competitor sits at 40-50% for a year. Maintaining competitiveness required roughly $20 billion over a decade.
The pivot was right. Abandoning leading-edge to focus on mature nodes (12nm and above) was the clearest moment of strategic lucidity. GlobalFoundries went public in 2021 at ~$30 billion — profitable specialty foundry with growing geopolitical relevance.
What hardware teaches software people
Software people think in sprints where failure costs hours. Hardware people think in $20 billion commitments where "fail fast" means wasting the GDP of a small country.
Reversible decisions — features, markets, pricing — deserve the software model: speed, experimentation, iteration. Irreversible decisions — which fab to build, which process node — deserve the hardware model: analysis, consensus, conviction. Most technology decisions are mixed. Knowing which model applies to which part of the decision is the leverage. (For deeper reading on where semiconductors and AI intersect today, Dylan Patel's SemiAnalysis is the best source.)