Anthropic has locked in multi-year chip and compute deals with Alphabet’s Google and Broadcom that could be worth hundreds of billions of dollars, the Financial Times reported, as the AI startup’s annualized revenues climb to about $30 billion. The scale reorders the AI supply chain overnight, elevating GOOGL and AVGO while raising fresh questions about Nvidia’s grip on the market and how hyperscalers intend to ration scarce compute.
The headline forces investors to redraw the AI leaderboard. Google (GOOGL) gets a clearer path to fill its custom TPU capacity and boost Google Cloud utilization. Broadcom (AVGO) cements its role as the go-to builder of custom accelerators and AI networking, converting long lead times into multi-year revenue visibility. Nvidia (NVDA), which still sells every H100 and B100 it can make, faces a louder message: the biggest buyers want alternatives on cost and control. AMD (AMD) sits somewhere in the middle, fighting for sockets where CUDA lock-in weakens and where customers demand price-performance beyond Nvidia’s premium. Down the stack, Taiwan Semiconductor (TSM) and HBM memory suppliers Micron (MU), SK Hynix, and Samsung remain the non-discretionary winners because every path to AI scale runs through their bottlenecks.
For Anthropic, the deals look like insurance against compute scarcity, not a vanity arms race. Expect take-or-pay structures, large prepayments, and capacity reservations bundled with cloud commitments. For Alphabet, this is about monetizing internally designed silicon at industrial scale, keeping workloads on TPUs, and pushing more inference to custom chips where unit economics can beat Nvidia’s general-purpose accelerators. Broadcom’s role is to deliver tailored ASICs and the plumbing that makes them sing: high-radix switches, co-packaged optics, and advanced packaging that squeezes more bandwidth per watt. None of this happens without TSMC’s CoWoS capacity and a steady pipeline of HBM3E and HBM4. The practical upshot: Anthropic offloads the capex shock to its suppliers and cloud partners while locking in a floor of compute it can sell into with premium-priced models.
A $30 billion annualized revenue figure reframes the argument about AI “hype.” You do not book that kind of run-rate without strong usage, sticky enterprise contracts, or a viable plan to monetize inference at scale. The more telling metric, though, is compute intensity per dollar of revenue. If Anthropic’s model mix is pushing toward heavier inference loads—agents that run longer, retrieval that calls multiple models, or safety stacks that double the pass count—then buying ahead on capacity protects gross margins. It also hints that unit costs are falling fast enough to support broader pricing without collapsing ARPU. This is how hyperscalers want the flywheel to work: commit to chips, drive utilization, compress costs, expand the market, then commit again.
Nvidia’s dominance has been built on time-to-market, CUDA software gravity, and a networking stack that works out of the box. What shifts when buyers like Anthropic hard-commit to alternatives is not just share, but the blended cost of goods sold for the entire AI sector. If Google and Broadcom can deliver comparable throughput at lower total cost of ownership—counting networking, memory, and developer time—Nvidia’s pricing power gets tested in training first, inference next. That does not mean NVDA loses near term; the backlog and HBM constraints still support tight supply and premium pricing. It does mean the next round of negotiations will lean harder on price-per-token, model latency, and energy efficiency. Watch for more heterogeneous fleets: TPUs for training certain architectures, AVGO-built inference ASICs for hot endpoints, NVDA where flexibility and mature tooling still win.
Anthropic has juggled relationships across Amazon (AMZN), Google (GOOGL), and other providers. Amazon’s investment and Bedrock integration made AWS a natural home. Today’s report suggests Google has secured a larger, longer runway of Anthropic workloads. If Broadcom is building dedicated silicon for Anthropic or for Google’s platform that Anthropic will rely on, expect more questions from regulators about exclusivity, bundling, and whether hyperscalers are using infrastructure control to steer AI market structure. The defense will be simple: nobody can meet AI demand without deep coordination across chips, fabs, packaging, and cloud. The legal risk rises if compute credits, financing, and capacity access become de facto lock-in tools that limit switching. Enterprises buying AI services will care less about antitrust theory than service-level guarantees and price stability—but governments are not going to ignore a compute cartel if one forms in practice.
For AVGO, this is textbook Hock Tan: secure multi-year, low-volatility cash flows, monetize NRE up front, and target niches where customization yields durable moat. Custom accelerators and AI networking sit squarely in that playbook. Broadcom already anchors Google’s TPU supply chain and dominates high-end Ethernet switching; co-packaged optics and advanced interconnect raise the switching costs further. The risk is execution—packaging capacity, HBM availability, and the physics of moving terabytes per second through a data center without melting the power budget. But if AVGO can keep schedule fidelity while pushing 51.2T and beyond in its switch silicon, the operating leverage is significant. That helps fund more R&D and keeps Broadcom present in any conversation about decoupling from Nvidia without taking untenable performance haircuts.
Capacity is the currency. TSMC’s CoWoS expansions, SK Hynix’s HBM ramp, and Micron’s HBM share gains are the gating factors for every player, including GOOGL and AVGO. Alphabet’s capex guide and commentary on TPU utilization will be immediate tells. Broadcom’s next print should break out AI more granularly—custom AI, switching, optics—and any change in backlog duration will matter. Anthropic’s product cadence will show whether the compute shows up as bigger models, cheaper tokens, or differentiated features for enterprises. On the competitive front, watch OpenAI’s and xAI’s infrastructure disclosures; Elon Musk has leaned into building his own compute footprint, and any pivot toward custom silicon there would reinforce the de-Nvidia narrative. AMD’s MI300 adoption tempo is the swing factor for a true three-horse race at the accelerator layer.
This is also a hyperscaler margin story. Alphabet benefits twice if Anthropic leans into TPUs on Google Cloud: higher utilization and lower unit costs versus third-party accelerators. Amazon faces pressure to respond with fresh commitments, discounts, or custom silicon acceleration through AWS Trainium and Inferentia to keep Anthropic workloads on Bedrock or win incremental AI tenants. Microsoft (MSFT), already tied at the hip with OpenAI and pushing its Maia accelerators, will likely match speed for speed to maintain Azure AI momentum. For all three, the KPI is the same: AI gross margin convergence with core cloud margins. Whoever gets closest to that steady state while holding or gaining share will be priced like the long-run winner.
If the FT’s reporting holds, this is another turn of the flywheel toward vertically integrated AI stacks led by hyperscalers and enabled by custom silicon. Near term, it supports GOOGL’s AI capex narrative and strengthens AVGO’s multi-year visibility. It nudges NVDA from absolute to relative winner as customers carve out tailored alternatives, though scarcity still props up Nvidia’s economics. The second-order beneficiaries remain TSM, MU, SK Hynix, Samsung, and the semicap complex that feeds them. The risk is execution: packaging, HBM supply, and software maturity can slip. The signal is clear: compute access, not model novelty, is the competitive moat in 2026. Cash is being converted into capacity at unprecedented scale, and the players who lock it down—on their terms—will set the price of AI for everyone else.
