Fact: On 23 September 2024, Goldman Sachs raised its twelve-month price target on Microsoft to $640 against a $501.61 close. Implied upside: 27.6%.
Fact: The note did not lead with Copilot seats. It led with silicon. MAIA โ Microsoft's in-house AI accelerator โ was named among the key areas of "value discovery."
Fact: For the quarter ending 30 June 2024, Microsoft reported capital expenditure including finance leases of $19.0 billion. Year over year, that is +77%. Full-year FY2024 capex approached $69 billion. Guidance for FY2025 clears $80 billion.
Fact: MAIA 100 sits on TSMC's 5-nanometer node. 105 billion transistors. 192GB of HBM3. Roughly 800 TFLOPS at FP8. An estimated 500-watt envelope.
Four inputs. One output. If the second-largest cloud operator on earth can push its own inference silicon into production, then the organizing premise of the decentralized compute sector โ that hyperscaler hardware margins constitute a permanent arbitrage for permissionless supply โ requires re-derivation from first principles. The sector has, by and large, not re-derived it. It is still pricing the 2021 curve.
I restate the numbers because they are the only inputs that matter here. Everything downstream is arithmetic. And arithmetic does not care how many Discord members a project has.
The AI-crypto convergence trade rests on a syllogism. GPUs are scarce. Hyperscalers capture 70โ90% gross margins on rented compute. Therefore idle consumer silicon and long-tail datacenter capacity can be aggregated permissionlessly, sold at a discount, and settled on a token that pays suppliers above their opportunity cost.
Each clause requires stress-testing, and each has been tested, publicly, by the companies the trade claims to disrupt. Scarcity is real but reallocating. Omdia's tracking puts Nvidia above 90% share of the datacenter AI accelerator market in 2023. That is a monopoly rent. It is also precisely the condition that guarantees the rent gets competed away โ by the buyers who pay it. Microsoft, Google, and Amazon are not passive victims of Nvidia pricing. They are the three largest purchasers of Nvidia silicon and the three largest designers of Nvidia substitutes. Google shipped TPU v1 in 2015 and is five generations deep. Amazon's Trainium arrived in 2020, with Trainium2 scheduled for EC2 UltraServer deployment and a multi-hundred-thousand-chip cluster in coordination with Anthropic. Microsoft is the last arrival with MAIA 100, announced at Ignite in November 2023, already validated internally on Bing and Office AI workloads, already racked inside Azure datacenters.
When your three largest customers all fund competing silicon, your pricing power has a half-life. The market understands this. It simply has not decided how long the half-life is. In the interim, token networks priced for infinity have been priced for a window.
Now take the mechanics apart, because the sector's public materials treat the stack as a black box and the box is where the margin lives.
A decentralized compute network is five components. There is a supply side โ heterogeneous GPUs, ranging from consumer cards to retired enterprise parts. There is a scheduler that receives jobs and assigns them, ideally by a reputation or stake-weighted routing rule. There is a telemetry and attestation layer that reports what actually executed, on what hardware, at what utilization. There is a settlement layer that pays suppliers in token, often with a vesting or challenge period. And there is a governance shell that owns the upgrade rights to the smart contracts.
Strip the branding and you have made a market for compute. Making a market is not the innovation. Making a market that clears cost-per-token below a hyperscaler's marginal cost is the innovation. That is the claim. That is the claim I want to audit.
The first unpriced dependency is the telemetry feed. Every one of these networks depends on a stream of measurements โ GPU model, memory bandwidth, uptime, job completion, throughput โ arriving from hardware the network does not own and cannot physically inspect. That feed determines pricing, reputation, and slashing. In DeFi I spent the better part of 2020 modeling a structurally identical problem: the price oracle. I simulated Compound's liquidation mechanics against historical Ethereum block data and found an edge case in the feed's latency profile that could drain collateral during high-volatility blocks. I wrote forty pages. The team called it theoretical. The mechanism was not theoretical. It was arithmetic applied to an input nobody controlled. That report established my baseline methodology: assume every external input is hostile until proven otherwise.
The compute sector has never produced an equivalent report for its own telemetry. Suppliers are economically motivated to over-report. The scheduler cannot distinguish a throttled 4090 from a healthy one without trusting a reported metric. Attestation via trusted execution environments reduces, in the end, to trusting a hardware vendor's root of trust โ a centralized dependency wearing decentralized clothing. And the slashing penalty that is supposed to deter misreporting is denominated in the same token whose price depends on the network appearing healthy. The incentive is not a deterrent. It is a circularity.
Protocol integrity is binary; trust is a variable. Either the job ran on the claimed silicon or it did not. There is no partial credit, and no amount of staking changes the round-trip latency between Frankfurt and Virginia.
The second unpriced dependency is verification, and it is where the discount dies.
Verifiable inference requires one of three mechanisms. Redundant execution: run the job two or three times on independent nodes and compare outputs. Cryptographic proof of the computation: SNARKs or similar over transformer inference. Or hardware attestation. Each has a price, and the price is not zero.
Redundant execution multiplies cost by the replication factor. If you run every job twice, you have halved your effective capacity. If you run it three times, you have cut it to a third. The discount you advertised against Azure was, in many published comparisons, computed on a per-GPU-hour basis that silently excludes this multiplier. Add it back and the spread compresses toward zero for anything but the smallest models.
Cryptographic proofs for general transformer inference remain orders of magnitude more expensive than the inference itself for realistic model sizes. That is not an engineering opinion. It is a published cost ratio, and it has not inverted.
Hardware attestation imports a vendor root of trust โ a single point of verification failure that no governance vote can decentralize, because the signing keys physically originate in a fab.
Now compare cost structures at the operator level. Microsoft owns the datacenter, the silicon, the hypervisor, the billing relationship, and the customer contract. Verification for Microsoft is an internal audit line on a P&L that already exists. Verification for a token network is the product's entire economic viability, funded from a treasury, priced by a market that has no visibility into the cost basis.
That is the asymmetry. It is not that decentralized compute cannot work. It is that the version that works is narrow, and the tokens are priced for the broad version.
I have executed this shape of analysis before, at a different layer. In 2023 I traced $4.3 billion in unbacked USDC movements from FTX to Alameda across wallet clusters and custody boundaries that the initial regulatory review had missed. The finding was never that transfers happened. The finding was that the accounting controls required to prevent them did not exist, and nobody had priced their absence. Redundant execution is the missing accounting control of decentralized compute. Nobody has budgeted for it, and every token in the sector trades as if it were free.
The third unpriced dependency is the cost peg, and it is the one that MAIA specifically attacks.
A decentralized network pays suppliers a rate that must exceed their opportunity cost. For a consumer GPU, opportunity cost approaches electricity plus depreciation, so networks can clear low and look cheap. For enterprise accelerators in a datacenter, opportunity cost is set by the prevailing market rental rate, which is set by hyperscaler list pricing, which is set by Nvidia's ASP plus a margin, plus a services wrap.
The network's cost floor is anchored to the exact price it claims to undercut.
That is not a moat. It is a peg. And pegs hold only while the anchor holds.
If Microsoft deploys MAIA at scale for internal inference, Azure's marginal inference cost decouples from Nvidia's ASP. Microsoft does not need to sell MAIA. It does not need to book a chip gross margin. It needs only to run its own workloads โ Bing, Office, Copilot, Azure OpenAI โ on silicon it owns, amortized across a denominator no token network can construct.
Run the denominator. FY2025 capex guidance clears $80 billion, with the overwhelming majority AI-attached. That is the fixed-cost pool. Against it sits the largest enterprise software distribution channel in existence: over 400 million paid Microsoft 365 commercial seats, Azure, GitHub, LinkedIn, Windows. Inference calls at that volume amortize silicon design cost to a rounding error. Azure AI already reported more than 60,000 customers, up 60% year over year, with AI services contributing 8 percentage points to Azure's 29% growth.
A token network's denominator is its emission schedule. Which is to say, its denominator is its own dilution.
Volatility is the tax on uncertainty. Here the uncertainty is not price. It is cost basis. And a network whose cost basis is pegged to a competitor's list price has already sold its margin to that competitor, whether or not it has noticed.
There is a structural parallel between token-funded compute and UST in 2022, and it is not rhetorical.
In early 2022 I tracked Terra's peg maintenance costs against LUNA sell pressure, built a Python model of daily burn rate, and called the decoupling roughly three weeks before it occurred. The mechanism was not sentiment. It was a subsidy that required continuous marginal demand to sustain a price relationship. When the marginal buyer stopped, the relationship broke โ and no community vote, no validator set, and no governance proposal could reconstruct it, because the break was arithmetic.
Token compute runs a lighter version of the same equation. Emissions pay suppliers above market for capacity. That capacity is then resold near or below its fully loaded cost, because the network is not charging for verification, redundancy, capital cost of hardware, or slippage on the token used to pay. The gap is funded by dilution. Usage growth is real. Unit economics growth is not. The two are routinely reported as though they were the same number.
The Terra comparison has limits, and I will state them. Decentralized compute is not promising a fixed peg. It is promising a discount. But the funding mechanism is identical: a token whose marginal buyer determines whether the subsidy continues. That is a demand curve masquerading as a protocol.
Recovery is not a phase; it is a reconstruction. Networks that survive a token drawdown do so by rebuilding their cost structure. Most will not, because rebuilding the cost structure requires admitting the discount was never real, and admitting that requires a governance vote that the treasury cannot survive.
Here is where governance becomes load-bearing. Every claim of decentralization in this sector routes through an upgrade key.
Compute networks are upgradeable smart contracts. The upgrade authority is a multisig, or a timelock owned by a small council, or a proxy admin held by the founding entity. The marketing language says protocol. The deployment script says four-of-seven.
This matters because cost structure is a governance variable. The emission schedule, the slashing parameters, the verification multiplier, the fee curve โ all of these are upgradeable, and the entity that can upgrade them sets the economics. A protocol that can adjust emissions to defend a token price can also adjust them to defend a treasury. The distinction between a protocol and a company is a function selector.
I have audited this exact failure mode in a different guise. In 2024, following the spot Bitcoin ETF approvals, I was contracted to review custody arrangements at three asset managers. One firm's multi-signature setup lacked proper key sharding, in direct contradiction of its own whitepaper language about institutional-grade security. I notified compliance, and the configuration was patched before public launch. The vulnerability was not cryptographic. It was procedural. Custody was decentralized in the document and concentrated in the operations runbook. Compute networks have the same document-versus-runbook gap, and it is wider, because there is no regulator demanding a patched runbook.
Code is law, but logic is the jury. The code upgradeable by four people is not law. It is a policy that happens to be executable.
The fourth structural problem is fragmentation, and it is the one the sector will not discuss because it implicates the token model itself.
There are now dozens of decentralized compute networks. They draw from overlapping supplier pools, chase the same enterprise buyers, list on the same exchanges, and depend on the same handful of market makers for liquidity. This is not scaling. It is slicing an already scarce addressable market into fragments, each with its own token, its own emission schedule, and its own bid-ask spread.
I have made this argument about Layer 2 rollups and it applies with greater force here, because compute networks lack even the shared settlement layer that rollups inherit from Ethereum. There is no composability ladder. There is no shared security budget. There is a scheduler, a token, and a slide deck.
Fragmentation has a specific mechanical cost. Thin books mean that any supplier seeking to convert earned tokens to fiat accepts meaningful slippage. That slippage is a real cost of production. It is paid in the same currency as the subsidy, and it is invisible in the headline rate the network advertises. When you see a network quoting a rate below Azure, ask what the supplier nets after conversion. Then ask who absorbs the difference. The answer is the marginal buyer of the token, which is the definition of a subsidy funded by future holders.
The bulls are not wrong about everything, and dismissing them wholesale is how analysts become furniture.
They are right that hyperscaler capex growth is not infinitely sustainable. $80 billion a year against a mid-$200-billion revenue base is a bet that requires the AI revenue curve to inflect before depreciation catches the income statement. FY2024 Q4 capex rose 77% year over year while Azure grew 29%, with AI contributing 8 percentage points. That is a widening gap. If revenue does not accelerate, the capex line becomes the story, and the $60 billion buyback authorized in September 2024 โ alongside a dividend raise from $0.75 to $0.83 โ becomes the shock absorber rather than the accelerant. A bear case exists inside Microsoft's own cash flow statement.
They are right that sovereignty, latency, and edge inference are real markets. Not every workload belongs in a hyperscaler region. Jurisdictional data residency, sub-10ms inference, and cost-sensitive batch work have legitimate reasons to sit outside Azure. Copilot+ PC, with 40+ TOPS NPUs across Qualcomm and Intel silicon and roughly 50 million Windows users granted access, is Microsoft conceding the same point: some inference belongs on the device, not in the region. A market that Microsoft concedes exists is a market.
And the bulls are right about UCIe. Standardized die-to-die interconnect and the open composable chiplet initiative โ Microsoft alongside AMD, Arm, and Samsung โ lower barriers for every smaller silicon designer over a multi-year horizon. That is a genuine tailwind. It is a tailwind for firms that design silicon. It is not a tailwind for token networks whose suppliers buy silicon retail.
Read that last distinction carefully, because it is the whole disagreement. The bulls have identified a real phenomenon โ the commoditization of accelerator design โ and attached it to the wrong instrument. The beneficiary of a commoditized design layer is the party with the highest volume of inference and the lowest cost of capital. That is not a distributed network of retail GPU owners. That is a hyperscaler with an $80 billion capex line and a 400-million-seat distribution channel.
There is a specific historical pattern worth naming, because I have watched it three times. In 2020, oracle latency was dismissed as theoretical. In 2022, the burn rate was dismissed as fearmongering. In 2023, the missing accounting controls were dismissed as an isolated failure at one exchange. Each time the mechanism was published in advance, with numbers, and each time the objection was that the mechanism was too boring to matter. Cost curves are boring. That is why they hold.
So here is what I am tracking, in order of evidentiary weight.
Azure AI gross margin commentary. Not Azure growth. Margin. If MAIA reaches scale and margin does not move, the vertical integration thesis is decorative. If margin moves, the correct conclusion is not bullish for the peer set. It is a downward re-rate, because a lower-cost incumbent compresses every supplier above it โ including every decentralized network whose cost basis is pegged to that supplier's list price.
The ratio of AI revenue contribution to capex growth. In FY2024 Q4 that ratio was 8 percentage points of contribution against 77% capex growth. That ratio has to invert within a handful of quarters, or the fixed-cost pool stops being an investment and becomes a liability. Watch the depreciation schedule, not the press release.
Third-party MAIA benchmarks. Not Microsoft's FP8 throughput slide. Sustained tokens per second per watt on a realistic model at realistic context length, measured by an entity with no position. Until that print exists, MAIA is a specification, and specifications have never paid a depreciation schedule. Same standard I applied to the ten AI-crypto projects whose IP ranges resolved to centralized cloud hosts: show me the log, or show me nothing.
The sector will not be killed by a price target. It will be killed by a cost curve that everyone could have computed and almost nobody did.
Here is the question I would put in writing to every decentralized compute treasury, with a response deadline attached: what is your all-in cost per million tokens for a 70-billion-parameter model at 32k context, including redundant execution for verification, including token conversion slippage, including hardware capital cost, and what is the emission-adjusted figure for the same job eight quarters forward?
If the answer is a deck, the answer is no. If the answer is a spreadsheet with a sensitivity table and a stated replication factor, there is a conversation.
The silicon race is not the story. The accounting beneath it is. And the accounting has been available the entire time โ which means the only remaining question is who chose not to read it.